Professional Documents
Culture Documents
Analysis and Optimization of Cyclone Separators Geometry Using RANS and LES Methodologies
Thesis submitted in fulllment of the requirements for the award of the degree of Doctor in de Ingenieurswetenschappen (Doctor in Engineering) by
Khairy Elsayed
Brussels, October 2011 Advisor: Prof. Dr. Ir. Chris Lacor
Analysis and Optimization of Cyclone Separators Geometry Using RANS and LES Methodologies
by
Khairy Elsayed
Submitted to the Department of Mechanical Engineering, in partial fulllment of the requirements for the degree of Doctor in Engineering
Analysis and Optimization of Cyclone Separators Geometry Using RANS and LES Methodologies Khairy Elsayed Department of Mechanical Engineering, Vrije Universiteit Brussel Pleinlaan 2, B-1050 Brussels, Belgium Thesis submitted in partial fulllment of the requirements for the academic degree of Doctor in Engineering
Promoter: Prof. dr. ir. Chris Lacor Jury: Prof. dr. ir. Johan Deconinck, voorzitter Prof. dr. ir. Rik Pintelon, vice-voorzitter Prof. dr. ir. Gunther Steenackers, secretaris Prof. dr. ir. Gert Desmet Prof. dr. ir. Harry van den Akker (Delft University of Technology, Netherlands) Prof. dr. ir. Herman Deconinck (Von Karman Institute, Belgium)
2011 Khairy Elsayed 2011 Uitgeverij University Press Leegstraat 15 B-9060 Zelzate Tel +32 9 342 72 25 E-mail: info@universitypress.be www.universitypress.be Vrije Universiteit Brussel Faculteit Ingenieurswetenschappen Pleinlaan 2 1050 Brussel Contact: +32 (0)2 629 39 10 http://www.vub.ac.be/IR secr-dtw@ir.vub.ac.be ISBN 978-94-9069-594-1 All rights reserved. No parts of this book may be reproduced or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of the author.
Abstract
The gas-solids cyclone separator is an industrial equipment that has been widely used for more than a century. Due to its industrial relevance, a large number of experimental, theoretical and computational studies have been reported in the literature aimed at understanding and predicting the performance of cyclones in terms of pressure loss and collection efciency (cut-off diameter). The currently used mathematical models for the prediction of cyclone performance, however, exhibit limited accuracy and generality. Moreover, the cyclone performance can be calculated using the articial neural networks approach. An alternative approach is to simulate the gas-particle ow eld in a cyclone by computational uid dynamics (CFD). As a result of the recent progress of computational power and numerical techniques, CFD has been widely applied to industrial ow problems. The cyclone performance parameters are governed by many operational parameters (e.g., the gas ow rate and temperature) and geometrical parameters. This study focuses only on the effect of the geometrical parameters on the ow eld pattern and performance of the tangential inlet cyclone separators using three different approaches, the most robust mathematical models, articial neural networks and CFD approaches. The study was limited to reverse-ow gas-solids cyclone separators operating at low solids loading. The objective of this study is four-fold. First, to determine the most signicant factors affecting the cyclone performance based on the previous studies and statistical analysis of data using response surface methodology. Second, to study each (signicant) parameter separately to obtain more details about its effect on the ow eld pattern and the performance. Third, to obtain the most efcient cyclone design for minimum pressure drop (using the most robust mathematical models, articial neural networks and CFD dataset). Finally, to obtain the most efcient cyclone design for best performance (minimum pressure drop and minimum cut-off diameter) using multi-objective optimization techniques with two different optimization techniques (both the Nelder-Mead with desirability function and the genetic algorithms (NSGA-II)). The response surface methodology has been performed using dataset obtained from the Muschelknautz method of modeling (MM) to determine the most signicant parameters. Four geometrical factors have signicant effects on the cyclone performance viz., the vortex nder diameter, the ini
let width, the inlet height and the cyclone total height. There are strong interactions between the effect of inlet dimensions and the vortex nder diameter on the cyclone performance. The same investigation has been repeated using articial neural network approach based on the experimental pressure drop. A radial basis neural network (RBFNN) is developed and employed to model the pressure drop for cyclone separators. The neural network has been trained and tested by experimental data available in literature. The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone pressure drop. The analysis indicates the signicant effect of the vortex nder diameter and the vortex nder length, the inlet width and the total height. Furthermore, Four mathematical models (Muschelknautz method MM, Stairmand, Ramachandran and Shepherd and Lapple) have been tested against the experimental values. The residual error (the difference between the experimental value and the model value) of the MM model is the lowest. The numerical simulations of cyclone ow were carried out by solving the unsteady-state, three-dimensional Reynolds averaged Navier-Stokes (RANS) equations combined with a closure model for the turbulent stresses and the large eddy simulation approach. The modeling of the cyclonic ow by computational uid dynamics (CFD) simulation has been reported before in the literature. Using the experimental data available in literature, a generic assessment was carried out on a number of turbulence closure models. Only the Reynolds stress turbulence model (RSM) and large eddy simulation (LES) captured the cyclone ow eld best compared to the experimental measurements. The CFD model also predicted the collection efciency, where the particulate phase was treated in a Lagrangian framework by tracking a large number of particles of different size classes through the computational domain. The stochastic nature of the particle motion due to the uid turbulence was taken into account by a particle dispersion model. There was a reasonable agreement between the calculated and measured cut-off diameter for both the RSM and LES results. The effect of the cone tip-diameter on the ow eld and performance of cyclone separators was investigated because of the discrepancies and uncertainties in the literature about its inuence. Three cyclones with different cone tip diameters were studied using large eddy simulation (LES). The ow eld pattern has been simulated and analyzed with the aid of velocity components and static pressure contour plots. The obtained results demonstrate that the cone tip-diameter has an insignicant effect on the collection efciency (the cut-off diameter) and the pressure drop. The simii
ulation results agree well with the published experimental results and the mathematical models trend. The effect of the cyclone inlet dimensions on the performance and the ow eld pattern has been investigated computationally using the Reynolds stress turbulence model (RSM) for ve cyclone separators. The maximum tangential velocity in the cyclone decreases with increasing the cyclone inlet dimensions. Increasing the cyclone inlet dimensions decreases the pressure drop. The cyclone cut-off diameter increases with increasing cyclone inlet dimension. Consequently, the cyclone overall efciency decreases due to weakness of the vortex strength. The effect of changing the inlet width b is more signicant than the inlet height a, especially for the cut-off diameter. The optimum ratio of inlet width to inlet height b/a is from 0.5 to 0.7. The effect of the vortex nder dimensions (both the diameter and length) on the performance and ow eld pattern has been investigated computationally using the large eddy simulation (LES) for nine cyclone separators. The maximum tangential velocity in the cyclone decreases with increasing the vortex nder diameter. Whereas, a negligible change is noticed with increasing the vortex nder length. Increasing the vortex nder length makes a small change in both the static pressure, axial and tangential velocity proles. However, decreasing the vortex nder diameter gradually changes the axial velocity prole from the inverted W to the inverted V prole. Decreasing the cyclone vortex nder diameter increases the maximum tangential velocity. The maximum tangential velocity approaches asymptotically 1.589 times the inlet velocity when decreasing the vortex nder diameter. The Euler number (dimensionless pressure drop) decreases with increasing the vortex nder diameter. Increasing the vortex nder length slightly increases the Euler number. The Stokes number increases with increasing the vortex nder diameter and slightly increases as the vortex nder length is increased. The effect of the cyclone height (both the barrel and cone) on the performance and ow eld pattern has been investigated computationally for six cyclone separators. The maximum tangential velocity in the cyclone decreases with increasing the cyclone (barrel or cone) height. Increasing the barrel height, makes a small change in the axial velocity, whereas increasing the cone height changes it considerably. Increasing the cyclone (barrel or cone) height decreases both the pressure drop and the cut-off diameter. The changes in the performance beyond h/D = 1.8 are small at constant cone height, whereas the performance improvement stops after hc /D = 4.0 (Ht /D = 5.5) at constant barrel height where h is the barrel height, hc is the cone height, Ht is the total cyclone height and D is the barrel diameter. iii
The effect of changing the cone height on the ow pattern and performance is more signicant than that of the barrel height. The CFD model was used to predict the pressure drop and the collection efciency of a range of cyclone geometries based on Stairmands highefciency design. These predictions were used to obtain an algebraic equation that relates the performance of a cyclone to its design and a limited set of dimensionless quantities (Euler number and Stokes number). This approach towards predicting cyclone performance by varying many geometrical parameters has not been reported before. To obtain new optimized cyclone separators, several optimization studies have been conducted in this thesis. Both the response surface methodology (RSM) and the radial basis function neural network (RBFNN) have been used as meta-models. Three different sources of data have been used to t the second order polynomial in case of RSM and for training the RBFNN. These data come from analytical models, experimental measurements and CFD simulations. Two optimization techniques have been used to optimize the cyclone geometry for minimum pressure drop, namely, the NelderMead and the genetic algorithms techniques. To handle the bi-objective optimization problem (both the pressure drop and the cut-off diameter), two approaches have been applied, the desirability function and NSGA-II techniques. All the new optimized cyclones obtained either for single objective or for bi-objective problems exhibit better performance than the Stairmand design. Moreover, a new correlation between the Stokes number and the Euler number is obtained. The new correlation can be used to estimate the Stokes number if the Euler number is known.
iv
Acknowledgments
First and foremost, I would like to gratefully acknowledge the enthusiastic supervision of my promoter Prof. Chris Lacor who gave me the opportunity to do a PhD under his guidance. I particularity thank him for our weekly technical discussions, which had a major inuence on this thesis. I am indebted to him for showing great condence in me and always pushing me to achieve greater heights, as well as for granting me sufcient freedom to pursue my own ideas. I can say for sure that the past years at VUB have been the most productive days of my learning. I also thank all the members of my thesis committee: Prof. Johan Deconinck, Prof. Rik Pintelon, Prof. Gunther Steenackers, Prof. Gert Desmet, Prof. Harry van den Akker and Prof. Herman Deconinck, whose constructive criticism and valuable suggestions improved the quality of this dissertation. I warmly thank the IT support of our system administrator Alain Wery. His support is invaluable for the research at our department. I greatly appreciate him for his good mood and everlasting patience through the perpetual stream of requests and computer problems coming towards him. I am yet to meet someone who is so patient and always ready to help others. Thank you very much, Alain! The support of our secretary Jenny Dhaes started even before I arrived in Belgium. She was there for help, starting from lling down my admission forms in Dutch, to organizing my PhD defense. Thanks a lot Jenny. A word of thanks should also go to Birgit Buys and my Egyptian colleague Mahmoud El-kafafy who helped me in printing the draft version. I am pleased to acknowledge my colleagues, Ghader Ghorbaniasl. The many discussions on mathematics and physics I have had with Ghader were always fruitful. Santhosh Jayaraju and Kris van den Abeele gave v
me the template that was used for this thesis, and in doing so, saved a lot of much needed time for me. I also enjoyed the scientic discussion with them. In this regard, I should mention Willem Deconinck as well. I taught the students a basic techniques in computer simulation course with him. He made it fun to do so with his pleasant mood and sense of humor, even though I was under the pressure of writing my thesis at that time. I would like to thank Willem once more for proofreading of some part of my thesis. At my rst days at VUB during which we were ofce mates, I have shared many laughs and a lot of joy with Mahdi Zakyani. I am pleased to acknowledge my present and former colleagues Dean Vucinic, Matteo Parsani, Patryk Widera, Xiadong Wang, Vivek Agantori, Florian Krause and Dinesh Kumar. We have nice discussions and good fun at the coffee corner. In addition, I am sincerely thankful to Prof. Momtaz F. Sedrak, Prof. Ahmed F. Helal, Prof. Mohammed M. Abdelrahman, Prof. Mohammed Fatouh and Prof. Samira Elshereef who played a major role in my scientic career. I have learned a lot from them during my Master and Bachelor studies. I consider them as good examples for Egyptian professors. All my thanks are given to the Egyptian community at Belgium for advices, support and continuous encouragement. Special thanks are given to Omar Ellabban, Sameh Sorror, Romany Abskharon, Ehab Khatab and Wael Mohammed for the help and advice they gave me during my stay here at Brussel, especially at my rst days at Belgium. I cannot forget to give all thanks to the spirit of my late parents who I am indebted with all my life. Lastly, and most importantly, my utmost gratitude is reserved to my dear wife and my two sons Omar and Ahmed for their patience and encouragement.
vi
Jury Members
President Prof. dr. ir. Johan Deconinck Vrije Universiteit Brussel
Vice-President
Secretary
Internal Member
External members
Prof. dr. ir. Harry van den Akker Delft University of Technology
Promoter
vii
viii
Contents
1 Introduction 1.1 Overview of dust collectors . . . . . . . . . . . . . . 1.2 Cyclone separators: types and principals . . . . . 1.2.1 Advantages and disadvantages of cyclones 1.2.2 Applications . . . . . . . . . . . . . . . . . . 1.2.3 Principals of cyclonic separation . . . . . . 1.2.4 Factors affecting the cyclone performance . 1.3 Motivation of this work . . . . . . . . . . . . . . . . 1.4 Outline of the thesis . . . . . . . . . . . . . . . . . 2 Literature Review 2.1 Classication of study approaches . 2.2 Mathematical models . . . . . . . . . 2.3 Experimental methods . . . . . . . . 2.4 Computational uid dynamics (CFD) 2.5 Discrepancy in the previous studies 2.5.1 The cone tip diameter . . . . . 2.5.2 The dust outlet geometry . . . 2.5.3 The inlet dimensions . . . . . 2.5.4 The vortex nder dimensions 2.5.5 The cyclone heights . . . . . . 2.5.6 Previous optimization studies 2.6 Summary and research plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1 2 5 6 7 10 11 13 15 15 16 17 17 18 18 19 19 20 21 21 23 25 25 28 29 31 32 37
3 Governing Equations 3.1 Turbulence . . . . . . . . . . . . . . . . . . . . . . 3.2 The governing equations for the gas phase . . . 3.2.1 Reynolds averaged Navier Stokes (RANS) 3.2.2 Reynolds stress model (RSM) . . . . . . . 3.2.3 Large eddy simulation (LES) . . . . . . . 3.3 Discrete phase modeling . . . . . . . . . . . . . . ix
3.3.1 Governing equations for the particles . . . . . . . . . 3.3.2 Modeling the particle phase . . . . . . . . . . . . . . . 3.3.3 Stochastic trajectory approach . . . . . . . . . . . . . 4 Sensitivity Analysis of Geometrical Parameters 4.1 Sensitivity analysis . . . . . . . . . . . . . . . . . 4.1.1 Response surface methodology (RSM) . . 4.1.2 Design of experiment (DOE) . . . . . . . . 4.1.3 Analysis of response surfaces . . . . . . . 4.1.4 Conclusions . . . . . . . . . . . . . . . . . 4.2 The cone-tip diameter . . . . . . . . . . . . . . . . 4.2.1 Numerical simulation . . . . . . . . . . . 4.2.2 Results and discussion . . . . . . . . . . . 4.2.3 The ow pattern in the three cyclones . . 4.2.4 Comparison with mathematical models . 4.2.5 Conclusions . . . . . . . . . . . . . . . . . 4.3 The dust outlet geometry . . . . . . . . . . . . . . 4.3.1 Numerical simulation . . . . . . . . . . . 4.3.2 Results . . . . . . . . . . . . . . . . . . . . 4.3.3 Conclusions . . . . . . . . . . . . . . . . . 4.4 Closure . . . . . . . . . . . . . . . . . . . . . . . . 5 The Vortex Finder Dimensions 5.1 Introduction . . . . . . . . . . . . . . . . . . 5.2 Numerical settings . . . . . . . . . . . . . . 5.2.1 Conguration of the tested cyclones 5.2.2 Solver settings . . . . . . . . . . . . . 5.2.3 Boundary conditions . . . . . . . . . 5.2.4 Grid independency study . . . . . . . 5.3 Results and discussions . . . . . . . . . . . 5.3.1 The axial variation . . . . . . . . . . 5.3.2 The ow pattern . . . . . . . . . . . . 5.3.3 The radial variation . . . . . . . . . . 5.3.4 The cyclone performance . . . . . . . 5.4 Conclusions . . . . . . . . . . . . . . . . . . 6 The Inlet Dimensions 6.1 Introduction . . . . . . . . . . . . . . . . 6.2 Numerical settings . . . . . . . . . . . . 6.2.1 Conguration of the ve cyclones 6.2.2 Boundary conditions . . . . . . . 6.2.3 Selection of the time step . . . . . 6.2.4 CFD grid . . . . . . . . . . . . . . x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37 40 41 45 45 46 47 48 49 52 54 63 64 70 71 72 73 76 87 88
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 The axial variation of the ow properties 6.3.2 The ow pattern . . . . . . . . . . . . . . . 6.3.3 The cyclone performance . . . . . . . . . . 6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . 7 The Cyclone Height 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . 7.2 Numerical settings . . . . . . . . . . . . . . . . . 7.2.1 Conguration of the tested cyclones . . . 7.2.2 Boundary conditions . . . . . . . . . . . . 7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . 7.3.1 The axial variation . . . . . . . . . . . . . 7.3.2 The radial variation . . . . . . . . . . . . . 7.3.3 The ow pattern . . . . . . . . . . . . . . . 7.3.4 The performance . . . . . . . . . . . . . . 7.3.5 The cone height versus the barrel height 7.4 Conclusions . . . . . . . . . . . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
113 113 115 120 124 127 127 128 128 128 130 130 131 135 137 141 146 147 147 149 150 154 158 158 164 169 179 185 186 187 194 202 204 215 215 220 231 235 239
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
8 Optimization 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Single-objective using MM model . . . . . . . . . . . . . 8.2.1 CFD comparison between the two designs . . . . 8.2.2 Conclusions . . . . . . . . . . . . . . . . . . . . . 8.3 Single-objective using RBFNN . . . . . . . . . . . . . . 8.3.1 Radial basis function neural networks (RBFNN) 8.3.2 Evaluation of different mathematical models . . 8.3.3 Design of experiment (DOE) . . . . . . . . . . . . 8.3.4 CFD Comparison between the two designs . . . 8.3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . 8.4 Multi-objective optimization using GA . . . . . . . . . . 8.4.1 Articial neural network (ANN) approach . . . . 8.4.2 Single objective optimization . . . . . . . . . . . 8.4.3 Optimal cyclone design for best performance . . 8.4.4 Conclusions . . . . . . . . . . . . . . . . . . . . . 8.5 Multi-objective optimization using CFD data . . . . . . 8.5.1 Design variables and approaches . . . . . . . . . 8.5.2 The desirability function . . . . . . . . . . . . . . 8.5.3 Articial neural network (ANN) approach . . . . 8.5.4 Optimization Using Genetic Algorithms . . . . . 8.5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . xi
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
9 Conclusions and Future Directions 9.1 Conclusions . . . . . . . . . . . . . . . . . . . . 9.1.1 The most signicant geometrical factors 9.1.2 The impact of geometry . . . . . . . . . 9.1.3 Optimization . . . . . . . . . . . . . . . . 9.1.4 Multi-objective optimization . . . . . . . 9.2 Future Directions . . . . . . . . . . . . . . . . . A Mathematical models A.1 General assumptions . . . . . . . . . . . . . . . A.2 Barth model . . . . . . . . . . . . . . . . . . . . A.3 The Muschelknautz method of modeling (MM) A.4 Stairmand model for pressure drop . . . . . . . A.5 Purely empirical models for pressure drop . . . A.6 Iozia and Leith model for the cut-off diameter A.7 Rietema model for cut-off diameter . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
243 243 243 244 245 246 248 251 251 252 256 259 260 261 262 263 263 266 268 268 272 281
B Optimization Techniques B.1 Nelder-Mead . . . . . . . . . . . . . . . . . . . . . . . B.2 Genetic algorithms (GA) . . . . . . . . . . . . . . . . B.2.1 Description of the genetic algorithm process B.2.2 Genetic operators . . . . . . . . . . . . . . . . B.3 Multi-objective optimization . . . . . . . . . . . . . . Bibliography
xii
Nomenclature
Upper-Case Roman Ai AR Bc CD Cin D Dx Eu Fdrag Frx HCS Ht K K Li Le Ln P P Pr Qin R R2 Rb Rij Rep ReR Rm Rx inlet cross sectional area [m] total inside area of the cyclone [m2 ] cyclone cone-tip diameter [m] particle drag coefcient inlet dust concentration, [kg/m3 ] cyclone barrel diameter [m] cyclone vortex nder diameter [m] Euler number [-] drag force [N] Froude number (inertia force / gravitational force) [-] length of the control surface [m] cyclone total height [m] uctuating kinetic energy [m2 /s2 ] vortex nder entrance factor distance between the inlet section and cyclone center [m] distance between the outlet section and the barrel top [m] cyclone natural length [m] uctuating kinetic energy production [m2 /s3 ] mean pressure [N/m2 ] Prandtl number [-] gas volume ow rate [m3 /s] cyclone radius [m] coefcient of multiple determination dust outlet radius [m] Reynolds stress tensor [m2 /s2 ] Reynolds number based on the relative particle velocity [-] cyclone body Reynolds number [-] geometric mean radius [m] vortex nder radius [m] xiii
S Sg Stk50 V
vortex nder length [m] geometrical swirl number [-] Stokes number at the cut-off diameter [-] cyclone volume [m3 ]
Lower-Case Roman a b d50 dp f fair fr fsm gi h hc k ks m m p p t tres ui ui u i upi vx vzw x x50 xi cyclone inlet height [m] acceleration [m/s2 ] cyclone inlet width [m] cut-off diameter [m] particle diameter [m] total friction factor [-] gas friction factor [-] friction factor due to wall roughness [-] friction factor for smooth wall [-] acceleration due to gravity in i direction [m/s2 ] barrel height [m] cone height [m] turbulent kinetic energy [m2 /s2 ] wall relative roughness of the cyclone wall [m] mass [kg] dust mass ow rate [kg/s] static pressure [N/m2 ] ow physical time [s] ow average residence time [s] ow velocity component in i direction [m/s] mean velocity [m/s] uctuating velocity component in i direction [m/s] particle velocity in i direction [m/s] mean gas velocity through the vortex nder [m/s] wall axial velocity [m/s] particle diameter [ m] cut-off diameter [m] position [m]
Upper-Case Greek P Pbody Px lter width [m] pressure drop in the cyclone [N/m2 ] pressure drop in the cyclone body [N/m2 ] pressure drop in the vortex nder [N/m2 ]
xiv
Lower-Case Greek ij t t bulk p str ij ij gas moment-of-momentum ratio at inlet [-] ratio of inlet width to the cyclone radius [-] Kronecker delta [-] turbulence dissipation rate [m2 /s3 ] viscous dissipation [m2 /s3 ] dynamic viscosity [kg/(m s)] turbulent (eddy) viscosity [kg/(m s)] kinematic viscosity ( = /) [m2 /s] turbulent (eddy) kinematic viscosity (t = t /) [m2 /s] spatial separation [m] gas density [kg/m3 ] bulk density of the solid [kg/m3 ] particle density [kg/m3 ] bulk density of the strand layer at the walls [kg/m3 ] viscous stress tensor [N/m2 ] subgrid scale stress tensor [N/m2 ]
Subscripts
CS
g
in
p
w
at the control surface gas at the inlet surface particle properties angular (tangential component) close to the wall
Abbreviations ANOVA ANN CFD CFL DNS DOE DPM GA Analysis Of Variance Articial Neural Networks Computational Fluid Dynamics Courant-Friedrichs-Lewy number Direct Numerical Simulation Design Of Experiment Discrete Phase Modeling Genetic Algorithm
xv
GCI GEC LDA LES MM PIV RANS RBFNN RNG RSM RSM SGS
Grid Convergence Index Grade Efciency Curve Laser Doppler Anemometry Large Eddy Simulation Muschelknautz Method of modeling Particle Image Velocimetry Reynolds Averaged Navier-Stokes Radial Basis Function Neural Network Renormalization Group Response Surface Methodology Reynolds Stress turbulence Model Subgrid Scale Model
xvi
Chapter 1 Introduction
1.1 Overview of dust collectors
There are four principal types of industrial dust collectors [77] namely, inertial separators, fabric collectors, wet scrubbers and electrostatic precipitators. The selection of one type depends mainly on the particle size as shown in Fig. 1.1. The classication of different dust collectors is shown in Fig. 1.2. The inertial separators separate dust from gas streams using a combination of forces, such as centrifugal, gravitational, and inertial. These forces move the dust to an area where the forces exerted by the gas stream are minimal. The separated dust is moved by gravity into a hopper, where it is temporarily stored. The three primary types of inertial separators are settling chambers, bafe chambers, and centrifugal collectors (e.g., cyclone separator). A settling chamber consists of a large box installed in the ductwork. The sudden expansion at the chamber reduces the speed of the dust-lled airstream and heavier particles settle down. Settling chambers are simple in design and can be manufactured from almost any material. However, they are seldom used as primary dust collectors because of their large space requirements and low efciency. A practical use is as precleaners for more efcient collectors. Bafe chambers use a xed bafe plate that causes the conveying gas stream to make a sudden change of direction. Large-diameter particles do not follow the gas stream but continue into a dead air space and settle. Bafe chambers are used as precleaners for more efcient collectors. Fabric collectors are commonly known as baghouses. Fabric collectors use ltration to separate dust particulates from dusty gases. They are one 1
Chapter 1. Introduction
Settling chambers
Cyclone separator
Liquid scrubbers
10-3
10-2
10-1
102
103
Figure 1.1: Suitable methods for removing particles from a gas stream [128]
of the most efcient types of dust collectors available and can achieve a collection efciency of more than 99% for very ne particulates. Dust collectors that use liquid are commonly known as wet scrubbers. In these systems, the scrubbing liquid (usually water) comes into contact with a gas stream containing dust particles. The greater the contact of the gas and liquid streams, the higher the dust removal efciency. The electrostatic Precipitators use electrostatic forces to separate dust particles from exhaust gases. A number of high-voltage, direct-current discharge electrodes are placed between grounded collecting electrodes. The contaminated gases ow through the passage formed by the discharge and collecting electrodes. The airborne particles receive a negative charge as they pass through the ionized eld between the electrodes. These charged particles are then attracted to a grounded or positively charged electrode and adhere to it [77].
Chapter 1. Introduction
pipe while the gas phase reverses its axial direction of ow and exits out through the vortex nder (gas outlet tube) [77]. Figure 1.3 shows a typical cyclone separator. The cyclone separator is one of the most efcient and robust dust separators. Its robustness results from lack of moving parts and the ability to withstand harsh operating environments. Moreover, cyclones are well suited for high pressure and temperature applications. Centrifugal collectors use cyclonic action to separate dust particles from the gas stream. In a typical cyclone, the dust gas stream enters tangentially forcing the ow into a spiral movement. The centrifugal force created by the circular ow throws the dust particles toward the wall of the cyclone. After striking the wall, the particles fall into a hopper located underneath. The most common types of centrifugal, or inertial, collectors in use today are single-cyclone separators and multiple-cyclone separators (multiclone). Single-cyclone separators create a dual vortex to separate the dust from the gas. The main vortex spirals downward and carries most of the heavier particles. The inner vortex, created near the bottom of the 4
Some disadvantages of cyclones are [77]: low efciency for particle sizes below their cut-off diameter when operated under low solids-loading conditions. usually higher pressure loss than other separator types, including bag lters and low pressure drop scrubbers. subject to erosive wear and fouling if solids being processed are abrasive or sticky. can operate below expectations if not designed and operated properly. Although this problem, as well as the erosion and fouling problem mentioned above, is not unique to cyclones. 5
Chapter 1. Introduction
1.2.2
Applications
Due to the mentioned advantages, cyclones have found application in virtually every industry where there is a need to remove particles from a gas stream. Figure 1.4 presents some examples of cyclones industrial applications with wide range of sizes, locations and applications. Today, cyclone separators are found in: ship unloading installations power stations spray dryers uidized bed and reactor riser systems (such as catalytic crackers and cockers) synthetic detergent production units food processing plants crushing, separation, grinding and calcining operations in the mineral and chemical industries fossil and wood-waste red combustion units (normally upstream of 6
a wet scrubber, electrostatic precipitator or fabric lter) vacuum cleaning machines dust sampling equipment Cyclones have also been used to classify solids on the basis of their characteristic such as their mass, density, size, or shape. Because of their simple construction and high reliability, cyclones are also used very effectively to separate two-phase gas-liquid mixtures, such as the entrained droplets exiting a venturi scrubber or other types of scrubber. Other examples include the removal of water droplets from steam generators and coolers and oil-mist from the discharge of air compressors. Likewise, they have been widely applied in process machinery to remove entrained oil and hydrocarbon droplets generated from spraying, injection, distillation, or most any process that results in the production of entrained droplets or a two-phase mixture. They have even been used as inlet devices to prevent foaming in gravity separation drums [77].
Chapter 1. Introduction
Dx S h D b Ht a
hc
Bc
Figure 1.5: Sketches of a reverse-ow, cylinder-on-cone cyclone with a tangential inlet. The geometrical notation is indicated in the right sketch
1. 2. 3. 4. 5. 6. 7.
the body diameter (barrel diameter) D the total height of the cyclone (from roof to dust exit) Ht the vortex nder diameter Dx the vortex nder length (from the roof of the separation space) S the inlet height a the inlet width b the height of the conical section hc or the height of the cylindrical section h 8. the cone-tip diameter (dust exit diameter) Bc
1.2.3.1 Real vortex ow Swirling ow, or vortex ow, occurs in different types of equipment, such as cyclones, hydrocyclones, spray dryers and vortex burners [77]. Two basic types of swirling ows can be distinguished: 1. forced vortex ow, which is a swirling ow with the same tangential velocity distribution as a rotating solid body 2. free vortex ow, which is the way a frictionless uid would swirl. 8
Tangential velocity
Fo r
ced
Rankine vortex
vo r
tex
Free vort ex
The tangential velocity in such a swirl is such that the moment-ofmomentum of uid elements is the same at all radii. The tangential velocity distribution in a real swirling ow is intermediate between these two extremes. Now imagine rst that the swirling uid has an innite viscosity (behaves like a solid body). Hence, no shearing motion exists between uid layers at different radii. In this case, the uid elements at all radial positions are forced to have the same angular velocity which equals v /r where v is the tangential velocity. This is the forced vortex ow or solid-body rotation: v = r (1.1)
In the other extreme, if the swirling uid has no viscosity, the motion of a given uid element is not inuenced by the neighboring elements at smaller and larger radii. If in such a uid, we bring an element to a smaller radius, its tangential velocity will increase, since its moment-ofmomentum (mv r) will be conserved. Such a vortex is called a free or frictionless vortex. In such a ow, we have rv = C , with C a constant, so that: C (1.2) r This is the second basic swirl ow. A real swirling ow normally has a core of near solid-body rotation surrounded by a region of near loss-free rotation as sketched in Fig. 1.6. This is called a Rankine vortex. The ow and pressure distribution within cyclones is more easily understood if we make clear the relation between static and dynamic pressures; p and 1/2v 2 , respectively, with the density. The well-known Bernoulli v = 9
Chapter 1. Introduction
equation for steady ow of a frictionless, constant density uid, which can be derived from the Navier-Stokes equations, states that: 1 p + gh + v 2 = constant along a streamline 2 (1.3)
In this equation, we recognize the static and dynamic pressures (the latter is often called the velocity head) as the rst and third terms on the left-hand side. They have been divided by the uid density. This equation shows that static and dynamic pressures can be interchanged in the ow eld [77]. In areas where the velocity is high, the static pressure will be low and vice versa. It is especially important to appreciate this interdependence between static and dynamic pressure when dealing with swirling ows [77]. The left-hand side of Eq. 1.3 is sometimes called Bernoullis trinomial. The second term is unimportant relative to the two others when discussing gas cyclones, since the uid density is relatively low, and height differences not very large. In an actual ow situation, the uid is not frictionless. Frictional dissipation of mechanical energy will therefore cause Bernoullis trinomial to decrease in the ow direction, i.e. the trinomial is no longer constant, but decreases along a streamline. Frictionless ow is, nevertheless, a reasonably good approximation in the outer part of the swirl in a cyclone; Bernoullis trinomial does not change very much there [77].
1.2.4
Figure 1.7 indicates the possible factors affecting the cyclone performance and ow pattern. These factors can be sub-classied as follows: 1. Cyclone dimensions Cyclone diameter Inlet height Inlet width Vortex nder diameter Vortex nder length Cylinder height Cyclone total height Cone tip diameter
4. Other factors Wall roughness Shape of vortex nder Eccentricity of vortex nder
Chapter 1. Introduction
12
Figure 1.7: Cause and effect plot for cyclone separator
13
Chapter 1. Introduction
14
2.5.1
Very little information is available on the effects of changing the cone bottom (tip) diameter, which determines the cone shape if other cyclone dimensions are xed [184]. Regarding this effect, discrepancies and uncertainties exist in the literature. Bryant et al. [17] observed that if the vortex touched the cone wall, particle re-entrainment occurred and efciency decreased, so collection efciency will be lower for cyclones with a small 18
+ 2.33
(2.1)
implying proportionality with the square of the inlet area. On the other hand, Ramachandran et al. [139] proposed, Eu = 20 ab 2 Dx 19
S D H h Bc DD D 1/3
(2.2)
2.5.4
The vortex nder size is an especially important dimension, which signicantly affects the cyclone performance as its size plays a critical role in dening the ow eld inside the cyclone, including the pattern of the outer and inner spiral ows. Saltzman and Hochstrasser [151] studied the design and performance of miniature cyclones for repairable aerosol sampling, each with a different combination of three cyclone cone lengths and three gas outlet diameters. Iozia and Leith [84] optimized the cyclone design parameters, including the gas outlet diameter, to improve the cyclone performance using their optimization program. Kim and Lee [95] described how the ratio of the diameters of cyclone body D and the vortex nder Dx affected the collection efciency and pressure drop of cyclones, and proposed an energy-effective cyclone design. Moore and Mcfarland [111] also tested cyclones, with six different vortex nders, and concluded that the variation in the gas outlet diameter under the constraint of a constant cyclone Reynolds number produced a change in the aerodynamic particle cut-off diameter. Recently, Hoekstra [74] investigated the effect of gas outlet diameter on the velocity prole using 2-D axisymmetric simulations. Lim et al. [102] examined experimentally the effect of the vortex nder shape on the collection efciency at different ow rates but without any explanation on its effect of the ow eld pattern and velocity proles. Raou et al. [140] duplicated numerically the same study of Lim et al. 20
24
Figure 3.1: Schematic representation of the energy cascade [32, 35]. is the Kolmogorov length scale, lDI is the length scale dividing the dissipation and the inertial subrange, lEI is the length scale dividing the energy-containing range and the inertial subrange, l0 is the length scale and L is the characteristic length [35].
(the large scales are of the order of the ow geometry), and a characteristic velocity u which is on the order of the turbulence intensity. If l and u are the length and velocity scales of the largest eddy, the time scale is derived as, l (3.1) u The large energy containing eddies give away their kinetic energy to slightly smaller-scale eddies with which the large scales interact. The process of kinetic energy transfer continues in a similar fashion until the smallest scale eddies are reached, where the frictional forces become so large that the kinetic energy is converted into internal energy. This process of energy transfer and dissipation is referred to as the energy cascade process. The scales at which the dissipation () takes place are the smallest scales, and are also referred to as the Kolmogorov scales. They can be estimated from the large-scale properties as follows, = = u2 u3 = l 26 (3.2)
3.1. Turbulence
Since the processes of dissipation in the smallest scales are due to viscous forces, the properties of the smallest eddies can be estimated using the ow kinematic viscosity ( ) and the dissipation () itself. The length, velocity and time scales are given by:
l =
1/4
(3.3)
1/4
u = () =
(3.4) (3.5)
1/2
The turbulent length scale l is related to the wave number as = 2/l. The energy spectrum E () for a turbulent ow is as shown in Fig. 3.2. From dimensional analysis, the Kolmogorov -5/3 law characterizes the inertial subrange which is given by, E () = C 2/3 5/3 C is the Kolmogorov constant. 27 (3.6)
ui (x, t) represents the i-th component of the uid velocity at a point in space x and time t. p(x, t) is the static pressure. ij (x, t) are the viscous stresses. (x, t) is the uid density. E and H are the total energy and total enthalpy per unit mass. qi in Eq. 3.9 is the heat ux which is proportional to the temperature gradient. T qi = (3.10) xi where is the thermal conductivity. The Mach numbers associated with air ow in cyclone separators are very nominal, which allows the ow to be treated as incompressible. Furthermore, the air behaves as a Newtonian uid, in which case the viscous stresses are related to the incompressible uid motion using a property of uid, viscosity [86]. 1 ij = 2 sij skk ij 3 sij is the instantaneous strain rate tensor given by, 28 (3.11)
sij =
1 2
ui uj + xj xi
(3.12)
For incompressible ows, Eqs. 3.7 and 3.8 are simplied to the following form,
uj xj ui ui + uj t xj
= =
0 1 p 2 ui + xi xj xj
(3.13) (3.14)
In this thesis, the temperature effects are ignored and hence Eq. 3.9 is uncoupled from the continuity and momentum equations. The four main numerical procedures for solving the Navier-Stokes equations are the direct numerical simulation (DNS), the large eddy simulation (LES), the detached eddy simulation (DES) and the reynolds averaged navier stokes (RANS) approach. The most accurate approach is DNS where the whole range of spatial and temporal scales of turbulence are resolved. Since all the spatial scales, from the smallest dissipative Kolmogorov scales (l ) up to the energy containing integral length scales (l), are needed to be resolved by the computational mesh, the number of points required in one direction is of the order, N= l l (3.15)
The number of points required for a resolved DNS in three dimensions can be estimated as, 9/4 3 l ul = Re9/4 (3.16) N= l The number of grid points required for fully resolved DNS is enormously large, especially for high Reynolds number ows, and hence DNS is restricted to relatively low Reynolds number ows. DNS is generally used as a research tool for analyzing the mechanics of turbulence, such as turbulence production, energy cascade, energy dissipation, noise production, drag reduction, etc [86].
Overline is a shorthand for the time average and in case of RANS, Ui Ui and ui =0. The above technique of decomposing is referred to as Reynolds Decomposition. Inserting this decomposition into the instantaneous equations and time averaging results in the Reynolds averaged Navier-Stokes equations (RANS).
U j xj U i U i + Uj t xj
= =
0 1 P 2U i ui uj + xi xj xj xj
(3.17)
(3.18)
ui uj in the last term of Eq. 3.18 represents the correlation between uctuating velocities and is called the Reynolds stress tensor. All the effects of turbulent uid motion on the mean ow are lumped into this single term by the process of averaging [86]. This will enable great savings in terms of computational requirements. On the other hand, the process of averaging generates six new unknown variables. Now, in total there are ten unknowns (3-velocity, 1-pressure, 6-Reynolds stresses) and only four equations (1-continuity, 3 components of momentum equation). Hence, we need six equations to close this problem. This is referred to as the Closure problem. Based on the way we close the Reynolds stress tensor, there are two main categories, namely the eddy viscosity models and the Reynolds stress model. The Reynolds stress tensor resulting from time averaging of the NavierStokes equations is closed by replacing it with an eddy viscosity multiplied by velocity gradients. This is referred to as the Boussinesq assumption. U j U i + xj xi
u i u j = t
(3.19)
where t is the turbulent (eddy) kinematic viscosity. In order to make Eq. 3.19 valid upon contraction because of Eq. 3.17, it should be rewritten as, 30
u i u j = t
U i U j + xj xi
2 + ij k 3
(3.20)
where ij is the Kronecker delta, ij = 1 if i = j and ij = 0 if i = j . k is the turbulent kinetic energy given by, k= 1 uu 2 i i (3.21)
The eddy viscosity is treated as a scalar quantity and is determined using a turbulent velocity scale v and a length scale l, based on the dimensional analysis.
t vl
(3.22)
There are different types of eddy viscosity models (EVM) based on the way we close the eddy viscosity. Algebraic or zero equation EVMs normally use a geometric relation to compute the eddy viscosity. In one equation EVMs, one turbulence quantity is solved and a second turbulent quantity is obtained from algebraic expression. These two quantities are used to describe the eddy viscosity. In two equation EVM models the two turbulent quantities are solved to describe the eddy viscosity. The interested reader can refer to [24, 177, 183] for more details. In the Reynolds stress models (RSM), an equation is solved for each Reynolds stress component as well as one length scale determining equation. However, RSMs are computationally much more demanding when compared to EVMs.
where Rij = u i uj is the Reynolds stress tensor. The RSM turbulence model provides differential transport equations for evaluation of the turbulence
31
With P being the uctuating kinetic energy production. t is the turbulent (eddy) kinematic viscosity; and k = 1, C1 = 1.8, C2 = 0.6 are empirical constants The transport equation for the turbulence dissipation rate, , is given as [98]: = + Uj t xj xj In Eq. (3.26), K = + t Ui 2 C 1 Rij C 2 xj K xj K (3.26)
1 u u is the uctuating kinetic energy, and is the 2 i i turbulence dissipation rate. The values of constants are = 1.3, C 1 = 1.44 and C 2 = 1.92.
3.2.3
The large eddy simulation technique was developed based on an implication from Kolmogorovs theory of self-similarity that the large eddies of the ow are dependent on the geometry while the smaller scales are more universal [86]. Hence, the big three-dimensional eddies which are dictated by the geometry and boundary conditions of the ow involved are directly calculated (resolved) whereas the small eddies which tend to be more isotropic are modeled. Large eddy simulation (LES) treats the large eddies more exactly than the small ones. In LES, the large scales in space are calculated directly as illustrated in Fig. 3.3. The LES technique is based on a separation between large and small scales. A grid size rst has to be determined. Those scales that are of a characteristic size greater than the grid size are called large or resolved scales, and others are called small or subgrid scales. The subgrid scales are included by way of a model called the subgrid model [154]. An 32
3.2.3.1 LES equations In case of RANS, the instantaneous continuity and momentum equations (Eq. 3.7 and 3.8) are time averaged to obtain a steady form of the averaged equations (Eq. 3.17 and 3.18). In case of LES, instead of time-averaging, the instantaneous time-dependent equations are ltered. Filtering is a method that separates the resolvable scales from the subgrid scales. Filtering can be performed in either wave number space or the physical space. The lter cut-off should lie somewhere in the inertial range of the spectrum (Fig. 3.2). In nite volume methods, box lters are always used because the nite volume discretization itself implicitly provides the ltering operation. One of the earliest volume average box lters was given by Deardorff [33].
(X, t) = =
1 3
x0.5x x0.5x
y 0.5y y 0.5y
z 0.5z z 0.5z
(, t)ddd
(3.27) (3.28)
+ s
In the above equation, denotes the resolvable scale ltered variable and s denotes the sub-grid scale uctuation. is the lter width given by = (xy z )1/3 . Leonard [100] dened a generalized lter as a convolution integral which is given by,
(X, t) =
G(X ; ) (, t) d3 33
(3.29)
G(X ; ) d3 = 1
(3.30)
The lter function in terms of the volume average box lter (Eq. 3.27) can be written as, 1/3 , |x | < x/2 0, otherwise
G(X ; ) =
Finally, the decomposition of the ow into a ltered part and a sub-grid part looks like, Ui P Tij = = = Ui + u s i P + ps
s Tij + ij
The hat operator . in the above equations represents the ltering operation as opposed to the time-averaging in case of RANS. Moreover, contrary to RANS, where the average of uctuations is zero, in LES, Ui = Ui [150] and us = 0. Further details on the ltering methods can be found in [32, 97, 132, 183]. Inserting the above decomposition into the instantaneous equations results in the following ltered Navier-Stokes equations,
Uj xj Ui Ui + Uj t xj
= =
0
sgs ij 1 P 2 Ui + xi xj xj xj
(3.31) (3.32)
3.2.3.2 SGS modeling From the energy cascade, explained in the beginning of this chapter, it is apparent that the energy transfer occurs from the bigger scales to the 34
sgs ij = Ui Uj Ui Uj
(3.33)
By using the denition of ltering as given by Eq. 3.28 we can further work sgs out ij as,
sgs ij s Ui Uj (Ui + us i )(Uj + uj ) s s s us i uj + (Ui uj Uj ui ) + Ui Uj Ui Uj Reynolds Crossterm Leonard
= =
(3.34) (3.35)
sgs ij
Leonard [100] shows that the Leonard stresses can signicantly drain energy from the resolvable scales and they can be directly computed. On the other hand, Wilcox [183] mentions that Leonard stresses are of the same order of magnitude as the truncation error when a nite-difference scheme of second-order accuracy is used, and thus it is implicitly represented. The cross-term stresses are dispersive in nature and largely account for the backscatter effects. Modeling them with a purely dissipative model such as Smagorinsky would be in conict because of its dispersive nature [97]. In many applications, it is assumed that the Leonard and cross-term stresses can be neglected, and only the Reynolds stresses remain to be modeled. It is the same case in the present work [86]. The interested reader is referred to Sagaut [150] for the detailed review of various SGS models available in literature. Smagorinsky model One of the simplest SGS model is the Smagorinsky model [161]. The unknown subgrid-scale stresses are modeled employing the Boussinesq assumption as in the case of RANS. The subgrid-scale stress are related to 35
where Ls is the length-scale for the sub-grid scale and is given by Cs V 1/3 , where V is the computational cell volume. It is interesting to note that the length scale is now the lter width rather than the distance to the closest wall as in RANS. Cs is a constant which is taken to be 0.17. The only disadvantage of the Smagorinsky model is the constant Cs , which is not really a constant, but is ow dependent. It is found to vary between 0.065 [110] and 0.3 [89]. In the dynamic version, which was rst proposed by Germano et al. [60], Cs is dynamically computed based on the information provided by the resolved scales of motion. The specication of Ls as Cs V 1/3 is not justiable in the viscous wall region as it incorrectly leads to a nonzero turbulent shear-stress at the wall. In order to rectify this, Moin and Kim [110] use a Van Driest damping function to specify the length scale as, Ls = Cs V 1/3 1 exp y+ A+ (3.38)
where y + = u d/ is the non-dimensional distance from wall, u is the wall shear stress velocity, d is the distance to the nearest wall and A=25 is the Van Driest constant. The above-described SGS model is a standard version as dened in Smagorinsky [161]. The LES simulations in the present thesis are performed employing the Fluent ow solver. The Smagorinsky model implemented in Fluent deviates slightly from the standard version in the following ways [86], The length-scale for the sub-grid scale is computed as min(d, Cs V 1/3 ). is the von Karman constant (typically a value of 0.41 is used), d is the distance to the closest wall. d is indeed one of the rst mixing length models in the literature to handle the turbulent viscosity and was proposed by Prandtl [133]. Van Driest damping is basically an improved version of Prandtls mixing length model. Both the Prandtl and the Van Driest model are algebraic and from the zero-equation models category. 36
The constant Cs in Fluent is taken to be 0.1 instead of 0.17 as was originally proposed. The value of 0.17 for Cs was originally derived for homogeneous isotropic turbulence in the inertial subrange. However, this value was found to cause excessive damping of large-scale uctuations in transitional ows near solid boundaries, and has to be reduced in such regions [59]. A Cs value of around 0.1 has been found to yield the best results for a wide range of ows, and is the default value in Fluent. Dynamic Smagorinsky-Lilly model Germano et al. [60] and subsequently Lilly [101] conceived a procedure in which the Smagorinsky model constant Cs is dynamically computed based on the information provided by the resolved scales of motion [59]. The dynamic procedure thus obviates the need for users to specify the model constant Cs in advance. The Smagorinsky model constant is dynamically computed instead of given as an input to the solver, but clipped to zero or 0.23 if the calculated model constant is outside this range to avoid numerical instabilities [59]. The second advantage of the dynamic SmagorinskyLilly model over the Smagorinsky model is the treatment near the wall. In the dynamic Smagorinsky-Lilly model, a damping function for the eddy viscosity near the wall is not required, since the model constant goes to zero in the laminar region just near the wall [55, 113].
p/e
Particles enhance production Negligable effect on turbulence Particles enhance dissipation
102
0
10
10-5
Two-Way
Coupling
10-3
10-1
Four-Way
Coupling
p
Particles
Particles Fluid
Particles
Fluid
Particles
Dilute suspension
Dense suspension
Figure 3.4: Map for particle-turbulence modulation [44]. p is the ratio of particles volume to the volume occupied by particles and uids. p is the particle response time, p = p d2 /(18), where p is the particle density, d is the particle diameter, is the uid viscosity. e is the turnover time of large eddy (time scale= l/u) [44].
One-way coupling: The phenomenon of mutual mass, momentum and energy transfer between the phases is termed as coupling. Elghobashi [44] proposed a map of regimes of interactions between particles and uid turbulence as shown in Fig. 3.4. For values of dispersed-phase volume fraction less than 106 , particles have negligible effects on turbulence and this is termed as one-way coupling. The volume fraction of dust particles we are dealing with in the present thesis is much less than 106 and hence one-way coupling is assumed. In the second regime which lies between 106 103 , the existence of particles can augment the turbulence if the ratio of the particle response time to the turnover time of a large eddy is greater than unity, or can attenuate turbulence if the ratio is less than unity. This interaction is called two-way coupling. In the third regime where the volume fractions are greater than 103 , in addition to two-way coupling between particles and turbulence, particle collisions take place and hence this regime is termed as four-way coupling.
38
Incorporating all the above assumptions, the Lagrangian equations governing the particle motion can be written as [86]: dxp dt dup dt
= =
up Fd (u up ) + gx (p ) p
(3.39) (3.40)
xp is the particle position, gx is the gravitational force, and p are the density of the uid and the particle respectively. Generally, the particle moves with a different velocity than the uid at any given point. The difference in uid velocity (u) and the particle velocity (up ), termed as the slip velocity (u up ), leads to an unbalanced pressure distribution as well as viscous stresses on the particle surface which yields a resulting force called drag force. In Eq. 3.40, the term Fd (u up ) is the drag force per unit particle mass. Fd is given by [86]: Fd = 1 Cd Rep p 24 (3.41)
where p is the particle relaxation time given by, p = Laws of drag coefcient The drag coefcient Cd is a function of particle Reynolds number (Rep ). Various experimentally based empirical correlations for the drag coefcient based on Rep are available in the literature. The Reynolds number of the particle is dened as: Rep = dp |u u p | (3.43) p d2 p 18 (3.42)
In Fluent, the drag coefcient for spherical particles is calculated by using the correlations developed by Morsi and Alexander [112]. It is given by, Cd = a1 + a2 a3 + Rep Re2 p (3.44)
where a1 , a2 and a3 are constants that apply to smooth spherical particles in a stipulated range of Rep as given in Table 3.1. 39
Rep < 0 .1 0 .1 < 1 .0 1 < 10.0 10.0 < 100.0 100.0 < 1000.0 1000.0 < 5000.0 5000.0 < 10000.0 10000.0 < 50000.0
3.3.2
Coming to the fundamental mathematical modeling of two-phase ow, the two most widely used approaches are the Eulerian continuum approach and the Lagrangian trajectory approach. Eulerian continuum approach In an Eulerian approach, the particles are treated as a second uid which behaves like a continuum and the equations are developed for average properties of the particles. For example, the particle velocity is the average velocity over an averaging volume. This approach is most suitable when one requires a macroscopic eld description of dispersed phase properties such as pressure, mass ux, concentration, velocity and temperature. Eulerian approach is more suitable for simulating large-scale particle ow processes. However, this approach requires sophisticated modeling in order to describe the key effects and phenomena found in industrial processes [31, 86]. Lagrangian trajectory approach A Lagrangian approach is useful when the particle phase is so diluted that the description of particle behavior by continuum models is not feasible. The motion of a particle is expressed by ordinary differential equations in Lagrangian coordinates and are directly integrated to obtain individual tracks of particles [86]. To solve the Lagrangian-equation for a particular moving particle, the dynamic behavior of the gas phase (generally obtained by an Eulerian approach) and other particles surrounding this moving particle should be pre-determined. Since the particle velocity and the corresponding particle trajectory are calculated for each particle, this approach is more suitable to obtain the discrete nature of motion of particles. However, to obtain statistical averages with reasonable accuracy, a 40
= =
up 1 (u up ) + g p
(3.45) (3.46)
The instantaneous uid velocity u in the above equation is represented as the sum of the mean and uctuating velocity, u=U +u Assuming isotropic turbulence, we have, u 2 = v 2 = w 2 = 2 k 3 (3.48)
(3.47)
where k is the turbulent kinetic energy. Furthermore, it is assumed that the local velocity uctuations of the uid phase obey a Gaussian probability density distribution. Most stochastic models in practical use are de41
2 k 3
(3.49)
where is a random number drawn from a normal probability distribution with zero mean and unit standard deviation. The minimal random number generator of Park and Miller with Bays-Durham shufe [135] is implemented [86]. The random number generator returns a uniform random derivative with zero mean and unit standard deviation. The chosen uctuation is referred to a turbulent eddy whose size (length scale) and life-time (time scale) is known. Sommerfeld et al. [164] proposed the following relations for eddy parameters, k 2 k 3
= =
ct te
(3.50) (3.51)
Figure 3.5 shows a 2-D schematic representation of an eddy inside a rectangular domain. At any given particle position (xp , yp ), the eddy parameters are rst evaluated based on the local uid kinetic energy and dissipation rate. The particle position (xp , yp ) is assumed to be located at the center of this hypothetical eddy. It is accepted that each eddy has its own uctuation u , which remains constant until the particle leaves this eddy. The particle leaving an eddy is based on a certain interaction time of the particle with the eddy. Once this interaction time is reached while time integration of particle equations, the particle is assumed to have left the present eddy. Now, based on the new position of the particle, new eddy pa rameters are calculated and a new uctuation u is assigned to this eddy. This procedure may be repeated for as many interaction times as required for the particle to traverse the required distance. If a statistically significant number of particles are tracked in this way, the ensemble averaged behavior should represent the turbulent dispersion induced by the prevailing uid eld [67]. The interaction time is the minimum of two time scales, one being a typical turbulent eddy lifetime and the other the crossing-time of the particle in the eddy [67]. tint = min(te , tc ) 42 (3.52)
where p is the particle relaxation time, le the eddy length scale and |u up | the magnitude of slip velocity. In circumstances where le /(p |u up |) > 1, Eq. 3.53 has no solution. This can be interpreted as the particle trapped by an eddy, in which case tint = te [67]. The mentioned eddy interaction model is needed only for RANS simulation to take into account the effect of turbulence on the particle. In LES simulations, the effect of the resolved velocity uctuations on the particles is accounted for and there is no need for an eddy interaction model like in RANS [86]. In this thesis, the effect of the subgrid scale velocity uctuations on the particle dispersion is assumed negligible and hence not modeled. Figure 3.6 represents a ow chart demonstrating the steps involved in tracking one injected particle.
43
Start with the location of one particle injected from the inlet surface at a certain point
Compute the surrounding cells of the current control volume where the particle lies
Determine the eddy parameters (and uctuating velocity using interpolated ow variables and the random number generator in case of RANS)
Determine the integration time-step based on the cell size & eddy parameters
No
Is the particle still in the current cell?
No
Is the current cell a boundary cell?
Check the distance between particle position & the nearest boundary cell (wall or outlet)
No
Figure 3.6: Flow chart demonstrating the steps involved in tracking one injected particle [86]
44
a/D 0.5
b/D 0.2
Dx /D 0.5
Ht /D 4.0 45
h/D 1.5
S/D 0.5
Bc /D 0.375
4.1.1
The usual method of optimizing any experimental set-up is to adjust one parameter at a time, keeping all others constant, until the optimum working conditions are found. Adjusting one parameter at a time is necessarily time consuming, and may not reveal all interactions between the parameters. In order to fully describe the response and interactions of any complex system a multivariate parametric study must be conducted [30]. Since there are seven geometrical parameters to be investigated, the best technique is to perform this study using the response surface methodology (RSM). RSM is a powerful statistical analysis technique which is well suited to model complex multivariate processes, in applications where a response is inuenced by several variables, and the objective is to optimize this response. Box and Wilson rst introduced the theory of RSM in 1951 [13]. RSM today is the most commonly used method of process optimization. Using RSM one may model and predict the effect of individual experimental parameters on a dened response output, as well as locating any interactions between the experimental parameters which otherwise may have been overlooked. RSM has been employed extensively in the eld of engineering and manufacturing, where many parameters are involved in the process [70, 106, 118, 167170]. 46
In order to conduct a RSM analysis, one must rst design the experiment, identify the experimental parameters to adjust, and dene the process response to be optimized. Once the experiment has been conducted and the recorded data tabulated, the RSM analysis software models the data and attempts to t a second-order polynomial to this data [30]. The generalized second-order polynomial model used in the response surface analysis was as follows:
7 7
Y = 0 +
i=1
i X i +
i=1
ii Xi2 +
i<j
ij Xi Xj
(4.1)
where 0 , i , ii , and ij are the regression coefcients for intercept, linear, quadratic and interaction terms, respectively. Xi and Xj are the independent variables, and Y is the response variable (Euler number).
47
4.1.3
For visualization of the calculated factor, main effects plot, Pareto chart and response surface plots were drawn. The slope of the main effect curve is proportional to the size of the effect, and the direction of the curve species a positive or negative inuence of the effect [61](Fig.4.2(a)). Based on the main effect plot, the most signicant factor on the Euler number are (1) the vortex nder diameter, with a second-order curve with a wide range of inverse relation and a narrow range of direct relation, (2) direct relation with inlet dimensions, (3) inverse relation with cyclone total height and insignicant effects for the other factors. Pareto charts were used to summarize graphically and display the relative importance of each parameter with respect to the Euler number. The Pareto chart shows all the linear and second-order effects of the parameters within the model and estimates the signicance of each with respect to maximizing the Euler number response. A Pareto chart displays a frequency histogram with the length of each bar proportional to each estimated standardized effect [30]. The vertical line on the Pareto chart judges, whether each effect is statistically signicant within the generated 48
response surface model; bars that extend beyond this line represent effects that are statistically signicant at a 95% condence level. Based on the Pareto chart (Fig. 4.2(b)) and ANOVA table (Table 4.3) there are four signicant parameters (six terms in the ANOVA table ) at a 95% condence level: the negative linear vortex nder diameter; the linear inlet width; the linear total cyclone height; a second-order vortex nder diameter; negative interaction between vortex nder diameter and inlet dimensions. These are the major terms in a polynomial t to the data. Therefore, the pareto chart is a perfect supplementation to the main effects plot. To visualize the effect of the independent variables on the dependent ones, surface response of the quadratic polynomial models were generated by varying two of the independent variables within the experimental range while holding the other factors at their central values [189]. Thus, Fig. 4.2(c) was generated by varying the inlet height and the inlet width while holding the other ve factors xed at their central value. The trend of the curve is linear, with more signicant effect for inlet width, with no interaction between the inlet height and width. The response surface plots given by Figs. 4.2(d), 4.2(e) and 4.2(f) show that there are interactions between both inlet width and inlet height with the vortex nder diameter. The effect of cyclone total height is less signicant with respect to the vortex nder diameter, but its effect is higher than that of the vortex nder length, the barrel height and the cone-tip diameter.
4.1.4 Conclusions
Mathematical modeling (the Muschelknautz method of modeling (MM)) has been used to understand the effect of the cyclone geometrical parameters on the cyclone performance. The most signicant geometrical parameters are: 1. 2. 3. 4. the vortex nder diameter the inlet section width the inlet section height the cyclone total height.
The effect of both the barrel height and the vortex nder length on the cyclone separator performance are small in comparison with these most signicant geometrical parameters. There are strong interactions between the effects of inlet dimensions and the vortex nder diameter on the cyclone performance. This study conrms the insignicant effect of the cone-tip diameter on the cyclone performance. However, the discrepancy exists in literature for this 49
Bold numbers indicate signicant factors as identied by the analysis of variance (ANOVA) at the 95% condence level.
50
(c) X1 versus X2
(d) X1 versus X3
(e) X2 versus X3
(f) X3 versus X4
51
al. [184] with the commercial nite volume code Fluent. Using different turbulence models they proved that Fluent with Reynolds stress turbulence model (RSM) predicts well the cyclone collection efciency and the pressure drop. The CFD simulation results from Chuah et al. [23] agree well with Xiangs experimental results in that cyclones with a smaller cone diameter result in a slightly higher collection efciency compared to cyclones with a bigger cone-tip diameter (only if the cone-tip diameter is not smaller than the gas exit tube diameter). Moreover, the change in the pressure drop will not be signicant when the cone size is varied. Both Xiang and Chuah did not give any results about the effect of the cone-tip diameter on the ow eld inside the cyclone separator, except some plots for axial and tangential velocity proles at two stations in the ow eld for Chuah et al. [23]. Xiang and Lee [185] computationally investigated the effect of the cone-tip diameter on the ow eld using the Reynolds stress turbulence model. They did not present any contour plots for either the static pressure, tangential and axial velocity. However, the comparisons between the tangential and axial velocity proles at different sections indicating no valuable difference between the three cyclones [185, Fig. 8, p. 216 and Fig. 9, p. 217 ], they mentioned that the cone-tip diameter has a signicant effect on the ow eld. No particle tracking study has been performed in the study of Xiang and Lee [185]. Currently a better understanding of the ow eld inside cyclone separators is an important concern, especially with the application of large eddy simulation (LES). The present study was undertaken in an effort to carry out a numerical study on the effect of the cone-tip diameter on the ow eld and the cyclone performance using LES available in Fluent commercial nite volume solver.
Table 4.4: The geometrical dimensions of the three cyclones
Dimension Body diameter, D Gas outlet diameter, Dx Inlet height, a Inlet width, b Cyclone height, Ht Cylinder height, h Gas outlet duct length, S Cone-tip diameter, Bc Cyclone I Cyclone II Cyclone III Length (mm) 31 15.5 12.5 5 77 31 15.5 19.4 15.5 11.6 Dimension ratio (dimension/D) 1 0.5 0.4 0.16 2.5 1 0.5 0.625 0.5 0.375
The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at a distance Li = 0.75D from the cyclone center (cf. Fig. 4.3).
53
Dx a h D
Le S Li b
Ht
Bc
Figure 4.3: Schematic diagram for the cyclone geometry and coordinate denition
4.2.1
Numerical simulation
4.2.1.1 Conguration of the three cyclones The cyclones used in this study had a reversed ow tangential inlet. The geometry and dimensions are shown in Fig. 8.1 and Table 4.4. Three cyclones with different cone-tip diameters are used viz., Bc /D= 0.625, 0.5 and 0.375. The three cyclones are identical to those used by both Xiang et al. [184, 186] and Chuah et al. [23]. Four plotting sections are used to investigate the effect of the cone-tip diameter Bc on the velocity proles as given by Table 4.5. 4.2.1.2 Selection of the turbulence model (RANS versus LES) For the turbulent ow in cyclones, the key to the success of CFD lies with the accurate description of the turbulent behavior of the ow [68]. To model the swirling turbulent ow in a cyclone separator, there are different turbulence models available in Fluent. These range from the standard k model to the more complicated Reynolds stress model (RSM)
Table 4.5: The position of different plotting sections
Section z (mm) z/D S1 5 0.16 S2 15 0.48 S3 30 0.97 S4 50 1.61
54
x is the time averaged x-velocity, v y is the time averaged y-velocity where v and is the angular coordinate. Boundary conditions and other settings Velocity inlet boundary condition is applied at inlet, outow at gas outlet and wall (no-slip) boundary condition at all other boundaries. The air inlet velocity Uin equals 8 m/s, corresponding to air inlet volume ow rate Qin =30 l/s, air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s, leading to a Reynolds number of 1.18E4 based on the cyclone diameter and the area averaged inlet velocity. The turbulence intensity I equals 5% and the turbulence characteristic length equals 0.07 times the inlet width [74]. At the cyclone inlet, the Reynolds stress specic method in Fluent solver is the Reynolds stress components. The diagonal components of the Reynolds 3 2 , I Uin stress tensor (normal stresses) are assigned to 2kin /3, kin = 2 where kin is the kinetic energy at the inlet [59, 81]. The shear stresses (non-diagonal components) at the inlet are set to zero. To take into account the stochastic component of the turbulent ow at the inlet for the LES simulation, articial perturbations have been generated using the spectral synthesizer method available in the Fluent solver [59, 82, 162], where the uctuation velocity components are computed by synthesizing a divergence-free velocity-vector eld from the summation of 100 Fourier harmonics [59]. The uctuations are added to the mean inlet velocity. The reason for introducing these articial perturbations instead of selecting the no-perturbation option in the Fluent solver, is that the unpertur57
N is the number of hexahedral cells, Eu is the Euler number (dimensionless pressure drop = pressure drop / average kinetic energy at inlet) and x50 is the cut-off diameter; the particle diameter that will produce 50% collection efciency (cf. Sec. 3.3). b The percentage absolute difference between the coarse and ne grid values for Euler number and cut-off diameter. c The percentage absolute difference between the medium and ne grid values for Euler number and cut-off diameter.
nomena like vortex-core precession (cf. Ref. [38]), the effect of the grid on the Strouhal number associated with the simulated vortex-core precession [38] should be included in the grid independency study probably requiring ner grids, but this is not part of the present study). Moreover, to evaluate accurately the numerical uncertainties in the computational results (especially because of the large difference between the results obtained on the coarse and the ne mesh which is about 7%), the concept of grid convergence index (GCI) was adopted using three grid levels per cyclone.
Grid convergence index (GCI) Roache [143145] suggested a quantitative measure for the grid convergence; the grid convergence index (GCI). The GCI can be computed using two levels of grid; however, three levels are recommended in order to estimate accurately the order of convergence and check that the solution is within the asymptotic range of convergence [160]. For a consistent numerical analysis the discretized equations will approach the solution of the actual equations as the grid resolution approaches zero [160]. The appropriate level of grid resolution is a signicant issue in numerical investigations. It is a function of many variables including the ow condition, type of analysis, geometry and many other variables. The GCI is based upon a grid renement error estimator derived from the theory of the generalized Richardson extrapolation [160]. The GCI is a measure of how far the computed value is away from the value of the asymptotic numerical value. Consequently, it indicates how much the solution would change with a further renement of the grid. A small value of GCI indicates that the computation is within the asymptotic range. 59
where Fs is a factor of safety. Fs = 3 for comparison of two grids and 1.25 for comparison over three grids or more. For the coarse grid: GCIcoarse = Fs ||rp (rp 1) (4.4)
is a relative error measure of the key variable f between the coarse and ne solutions, = f2 f1 f1 (4.5)
where f2 is the coarse-grid numerical solution obtained with grid spacing h2 . f1 is the ne-grid numerical solution obtained with grid spacing h1 . r is the grid renement ratio (r = h2/h1 > 1). For complicated geometries r is replaced by the ratio of the number of control volumes in the ne and coarse mesh [107] which is the case in this study, N1 N2
1 D
r12 =
(4.6)
where D = 2 and 3 for two-dimensional and three-dimensional geometries respectively [143, pp. 410]. N1 and N2 are the number of control volumes in the ne and coarse mesh respectively. p is the order of the discretization method. p equals two if the second order discretization is used for all terms in space [107] (However, Slater [160] stated that if all discretization in space was of second-order, p will be less than 2. The difference is due to grid stretching, grid quality, non linearity in the solution, presence of shocks, turbulence modeling and perhaps other factors). For the grid renement study, three meshes have been used with N1 , N2 and N3 cells for the ne, medium and coarse three-dimensional mesh respectively.
N2 1 r12 = N , r23 = N , e12 = f2 f1 , e23 = f3 f2 , where ei,i+1 = N2 3 fi+1 fi is the difference in the key variable f resulting from the use of different grids. If r12 = r23 then,
1 3 1 3
p = ln
e23 e12 60
/ln(r)
(4.7)
Equation 4.8 is transcendental in p. Using the iterative technique with relaxation factor introduced in Roache [144, 145] p = + (1 ) where = ln( ) ln(r12 ) (4.9)
p 1)e23 (r12 , = 0.5 and is the previous iteration of p. The p (r23 1)e12
author suggest to use = ln will stop if GCIne 23 = Now one can calculate, 12 =
1.25|23 | . p 1) (r23 | p p |
e23 e12
< 1E 5.
f2 f1 f1 ,
GCIne 12 =
1.25|12 | p 1) (r12
and
GCIne 23
(4.10)
The Richardson extrapolation can be used to obtain the value of f when the grid spacing h vanishes (h 0) [2, 143].
p fexact = f1 + (f1 f2 )/ (r12 1)
(4.11)
Table 4.7 presents the grid convergency calculations using GCI method and three grid levels for cyclones I -III. The following conclusions have been obtained from the GCI analysis: The results are in the asymptotic range for the three cyclones, because the obtained values for are close to unity. The ratio R is less than unity this means monotonic convergence [2] (Ali et al. [2] classied the possible convergence conditions into three groups, namely (1) monotonic convergence; 0 < R < 1 (2) oscillatory convergence; R < 0 (3) divergence; R > 1.) . There is a reduction in the GCI value for the successive grid renene ments (GCIne 12 < GCI23 ) for the two variables (Eu and x50 ). This indicates that the dependency of the numerical results on the cell size has been reduced. Moreover, a grid independent solution has been achieved. Further renement of the grid will not give much change 61
Table 4.7: Grid convergency calculations using GCI method and three grid levels for cyclones I - III
i 0c 1 Eu I x50 2 3 0 1 2 3 0 1 Eu II x50 2 3 0 1 2 3 0 1 Eu III x50 2 3 0 1 2 3
a b c
Ni 1021616 861077 632153 1021616 861077 632153 1025778 863852 513021 1025778 863852 513021 1027982 712576 513991 1027982 712576 513991
ri,i+1
ei,i+1
i,i+1
Ra
1.0586 1.1085 2.4800 1.3471 1.3500 1.0586 1.3550 1.1085 1.3960 2.2523 2.2700 1.0589 2.2800 1.1897 2.3600 1.2413 1.2500 1.0589 1.2570 1.1897 1.3350 2.6835 2.6870 1.1299 2.7120 1.1150 2.8690 1.1962 1.2000 1.1299 1.2100 1.1150 1.2400
0.0150 0.0750
0.0063 0.0312
0.2013 1.5597
1.0063
0.0050 0.0410
0.0037 0.0303
0.0100 0.0800
0.0044 0.0351
0.0070 0.0780
0.0056 0.0621
0.0250 0.1570
0.0093 0.0579
0.0100 0.0300
0.0083 0.0248
R=12 /23 . p GCI12 /GCI23 . = r12 The value at zero grid space (h 0). i=1, 2 and 3 denote the calculations at the ne, medium and coarse mesh respectively.
62
1.45
2.8
1.4
Cut-off diameter
Euler number
2.6
1.35
5E-07
1E-06
1.5E-06
1.15
(h --> 0)
-1
Figure 4.4: Qualitative representation of the grid independency study. The Euler number and the cut-off diameter for each cyclone at the three grid levels. N 1 is the reciprocal of the number of cells, h 0 means the value at zero grid size (cf. Table 4.7). To obtain a smooth curve; the spline curve tting has been applied in Tecplot post-processing software.
1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 LDA Simulation -0.5 0 0.5 LDA Simulation -0.5 0 0.5
0.4
0.2
-0.2
-0.4
Figure 4.5: Comparison of the time averaged axial and tangential velocity between LDA measurements [74] and the LES simulations at section Z=94.25 cm from the cyclone bottom (cf. Hoekstra [74] for more details about the used cyclone geometry in this validation).
4.2.3
The pressure eld Figure 4.7 shows the time-averaged static pressure contours plots. In the three cyclones, the static pressure decreases radially from the wall to the center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, while the gradient in the axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the low-pressure zone at the cyclone center (twisted cylinder). The static pressure contour plots for the three cyclones are almost the same. The velocity eld Based on the contours plots of the time averaged tangential velocity (Fig. 4.7) and the radial proles at sections S1, S2, S3 and S4 shown in Fig. 4.8 (cf. Table 4.5), the following comments can be drawn. The maximum tangential velocity equals around 1.25 times the average inlet velocity and occurs in the annulus cylindrical part. The tangential velocity distribution for the three cyclones are nearly identical at the corresponding sections. The tangential velocity prole at any section is composed from two regions, inner and outer. In the inner region the ow rotates approximately like a solid body rotation (forced vortex), where the tangential velocity increases with radius. After reaching its peak the tangential velocity decreases with radius in the outer part of the prole (free vortex). This prole is so-called 64
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.75
0.75
-0.5 -1
-0.75
0.75
Figure 4.6: Comparison of the time averaged tangential and axial velocity between RSM results of Xiang and Lee [186] and the current LES results at Z/D=1.29 for Cyclone III (Bc /D = 0.375). Note: the dimensionless radial distance = the distance / the cyclone radius, the dimensionless velocity = the velocity / inlet velocity.
Rankine type vortex which include a quasi-forced vortex in the central region and a quasi free vortex in the outer region. The radial proles given in Fig. 4.8 represent the time averaged tangential velocity in the lower part of the cyclone. The tangential velocity distributions at the bottom sections show good axis-symmetrical distribution. The tangential velocity proles for the three cyclones are identical in the inner region, where the maximum tangential velocity nearly equals the inlet velocity and occurs at a position 0.25 - 0.45 of the cyclone radius as given in Table 4.8. The axial velocity contours (Fig. 4.7) indicate the existence of two ow streams. Downward ow directed to the cyclone bottom (negative axial velocity), and upward ow directed to the vortex nder exit. The axial velocity plots for the three cyclones are nearly identical to the corresponding sections in the conical part. The axial velocity equals zero at the walls and maximum close to the position of maximum tangential velocity. The axial velocity proles shown in Fig. 4.8 exhibit a severe asymmetrical feature.
Table 4.8: Comparison between the maximum tangential velocity value and its position at different sections.
Section Cyclone / v v in max x/R I 1.08 0.3 S1 II 0.97 0.28 III 1.02 0.26 I 1.02 0.3 S2 II 0.95 0.33 III 1.05 0.33 I 0.96 0.32 S3 II 0.89 0.37 III 1.04 0.41 I 0.94 0.35 S4 II 0.88 0.43 III 1.035 0.425
The ratio between the maximum tangential velocity and the area average inlet velocity. The dimensionless distance between the centerline and the point of maximum velocity, R is the cyclone radius.
65
Comparison of the velocity proles in the three cyclones However, from the previous discussions it is clear that, the effect of the cone-tip diameter on the ow eld in the conical section is insignicant, in comparison with other geometrical parameters such as the vortex nder diameter. Nevertheless, in this section a comparison between the axial and tangential velocity proles at four sections (Table 4.5) will be analyzed as presented in Fig. 4.9. The tangential velocity proles in the forced vortex region are nearly identical in the three cyclones at each sections. The tangential velocity in the free vortex region increases as the cone-tip diameter is reduced. The tangential velocity proles for the three cyclones are almost the same. The axial velocity prole has the shape of an inverted W for all cyclones. The highest axial velocity occurs at 0.25 - 0.5 of the cyclone radius down the vortex nder until the cyclone bottom, and between 0.25 and 0.5 of the cyclone radius in the annulus and through the vortex nder. No considerable difference exists in the axial velocity proles for the three cyclones. Since the axial velocity proles is almost the same for the three cyclones, the average residence time of particles is nearly the same. Furthermore, the position of the highest axial velocity moves inward in the conical part as the cone-tip diameter reduced. From the previous analysis, the region of downward ow is nearly the same, for the three cyclones, while the tangential velocity slightly increases as the cone-tip diameter reduced, so the particles will experience a higher tangential velocity for cyclone III than in other cyclones for the same time (as the region of downward axial velocity is nearly equal). This results in a slightly higher collection efciency. This is consistent with the measured results reported by Xiang et al. [184] and simulation by Xiang and Lee [186] and Chuah et al. [23]. The change of the cone-tip diameter affects the ow eld in the cyclone separator but this change is so limited, i.e., the reduction of cone-tip diameter enhances the collection efciency but with a small percentage, as the ow eld pattern is so closed for the three cyclones. The DPM results In order to calculate the effect of the cone-tip diameter on the cut-off diameter, 104 particles were injected from the inlet surface with particle velocity equals the gas inlet velocity. The particle density is 860 kg/m3 and the maximum number of time steps for each injection was 9E5 steps. The DPM analysis results for the three cyclones are shown in Table 4.9 and Fig. 4.10. It is found that the cut-off diameter decreases slightly with decreasing the 66
Figure 4.7: The contours plots for the time averaged ow variables at Y=0 . From top to bottom: the static pressure N/m2 , the tangential velocity and axial velocity m/s. From left to right cyclone I, II and III respectively.
67
S1 S2 S3 S4
S1 S2 S3 S4
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.5 -1
-0.75
0.75
1.25
S1 S2 S3 S4
0.5
S1 S2 S3 S4
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.5 -1
-0.75
0.75
1.25
S1 S2 S3 S4
0.5
S1 S2 S3 S4
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.5 -1
-0.75
0.75
Figure 4.8: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) for each cyclone. From top to bottom: Cyclone I, II and III respectively.
68
0.5
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.75
-0.5 -0.25 0 0.25 0.5 Dimensionless distance Cyclone I Cyclone II Cyclone III
0.75
-0.5 -1
-0.75
-0.5 -0.25 0 0.25 0.5 Dimensionless distance Cyclone I Cyclone II Cyclone III
0.75
1.25
0.5
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.75
-0.5 -0.25 0 0.25 0.5 Dimensionless distance Cyclone I Cyclone II Cyclone III
0.75
-0.5 -1
-0.75
-0.5 -0.25 0 0.25 0.5 Dimensionless distance Cyclone I Cyclone II Cyclone III
0.75
1.25
0.5
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.75
-0.5 -0.25 0 0.25 0.5 Dimensionless distance Cyclone I Cyclone II Cyclone III
0.75
-0.5 -1
-0.75
-0.5 -0.25 0 0.25 0.5 Dimensionless distance Cyclone I Cyclone II Cyclone III
0.75
1.25
0.5
0.25
0.75
0.5
0.25
0 -1
-0.75
0.75
-0.25 -1
-0.75
0.75
Figure 4.9: Comparison between the radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0). From top to bottom: section S4 - S1 respectively.
69
cone-tip diameter while the pressure drop is increasing slightly. Consequently, the effect of the cone-tip diameter on the cyclone performance is insignicant. The trend of changing the cut-off diameter with the cone-tip diameter given by Chuah et al. [23] (Qin = 60l/min) supports the conclusion of the slightly decrease of the cut-off diameter by decreasing the cone-tip diameter (insignicant effect), Table 4.9.
4.2.4
Table 4.10 presents a comparison between the Euler number (dimensionless pressure drop) and the cut-off diameter obtained from CFD, experimental investigation [184] and seven mathematical models, viz. (the Barth model [9], the Muschelknautz method of modeling (MM) [116, 174], the Stairmand model [165], the Casal and Martnez-Benet model [21], the Shepherd and Lapple model [157], the Iozia and Leith model [85] and the Ritema model [142] (cf. Sec. 2.2) The Euler numbers obtained from the models of Shepherd and Lapple, Casal and Martnez-Benet are constant, because these models do not include the effect of the cone-tip diameter Bc in their formulas. The three other models (Barth, MM and Stairmand) indicate less effect on both the Euler number and the cut-off diameter by changing the cone-tip diameter. The models of Iozia and Leith in addition to that of Rietma indicate no change in the cut-off diameter with changing the cone-tip diameter. The results of mathematical models and the experimental investigation support the CFD results that the cone-tip diameter has an insignicant effect on the cyclone separator performance.
Table 4.9: The cut-off diameter and pressure drop for the three cyclones Cyclone Bc /D Cut-off diameter [m] Cut-off diameter [m] (Chuah et al. [23] ) Pressure drop [N/m2 ] Euler number Eu
Qin = 60 l/min.
70
1.8
3 1.4 2
1.2 1
0 0.3
0.35
0.4
0.45
0.5 Bc /D
0.55
0.6
0.65
1 0.7
Figure 4.10: The effect of cone-tip diameter on the pressure drop (Euler number) and the cut-off diameter (with spline curve tting to get a smooth curve).
4.2.5 Conclusions
Large eddy simulation has been used to study the effect of the cone-tip diameter on the cyclone ow eld and performance. Three cyclones with different values of Bc /D viz. 0.625, 0.5 and 0.375 (at constant vortex nder diameter Dx /D = 0.5) have been investigated. The following conclusion can be drawn. The cone-tip diameter has an insignicant effect on the ow pattern and performance. As the cone-tip diameter decreases, the maximum tangential velocity increases slightly, while its position is almost the same.
Table 4.10: The cyclone performance parameters using CFD, Experimental [184] and different seven mathematical models
Euler number Eu [-] Cyclone I II III I II III CFD 2.39 2.27 2.687 1.35 1.25 1.2 Barth 6.94 7.40 7.43 1.22 1.22 1.28 MM 4.88 4.95 4.95 2.13 2.089 2.089 Stairmand 6.68 6.69 6.69 Sphered 4.1 4.1 4.1 Casal 4.07 4.07 4.07 Iozia 1.44 1.44 1.44 Exp. 2.8 2.8 3.25 3.01 2.60 2.36
The mathematical model used for estimation of the pressure drop only. The mathematical model used for estimation of the cut-off diameter only. Different particle density.
71
Table 4.12 gives more details for the used cyclones, including the number of cells, cyclone volume and the ow residence time for each cyclone. Nine sections are used to plot the velocity proles as shown in Table 4.13.
(a) I
(b) II
(c) III
(d) IV
Figure 4.11: Schematic diagrams and surface meshes for the four tested cyclone separators
4.3.1.2 Solver settings Based on the study of Kaya and Karagoz [91] for the best selection of numerical schemes to be used with RSM model, the following discretization schemes have been used. The PRESTO scheme has been used for the pres74
The outlet section is above the cyclone surface by Le = 0.5D. The inlet section located at a distance Li = 0.75D from the cyclone center, the height of the dustbin and the dipleg, LD = 2D.
sure interpolation, the SIMPLEC algorithm for pressure velocity coupling, the QUICK scheme for momentum equations, the second-order upwind for the turbulent kinetic energy and the rst-order upwind discretization scheme for the Reynolds stresses [52], cf. Sec. 4.2.1.3 for more details. Chuah et al. [23] stated that the time step should be selected as a tiny fraction of the residence time tres . From Table 4.12, the value of tres varies between 0.237 and 0.399 s. Therefore a time step of 1E-4 is an acceptable value for the current simulations [52]. The simulations have been performed using FLUENT 6.3.26. 4.3.1.3 CFD grid and boundary conditions The mesh sensitivity study has been performed for the four tested cyclones with three levels for each cyclone, to be sure that the obtained results are grid independent. For example, three different meshes with respectively 130596, 260230 and 478980 cells have been used for cyclone I. The computational results on the three grids are presented in Table 4.14. As the maximum difference between the results is less than 5%, so the grid template 130596 produces the grid independent results [52, 146]. It has been observed that even 130596 grid provides a sufcient grid indepenTable 4.12: The details of the four tested cyclones
Cyclone Number of cellsa Cyclone volume x102 [m3 ] tres [s] b
a b
The total number of hexahedral cells after the mesh sensitivity study The average residence time, tres = V /Qin where V is the cyclone volume and Qin is the gas ow rate.
75
dency. However, for excluding any uncertainty, computations have been performed using the 260230 cells grid, where the total number of grid points was not that critical with respect to the computation overhead [10]. Figure 4.11(e) shows the surface mesh of the four cyclones. The hexahedral computational grids were generated using the GAMBIT grid generator. The boundary condition at the inlet section is the velocity inlet. An outow boundary condition is used at the outlet. The no-slip (wall) boundary condition is used at the other boundaries [52]. The air volume ow rate Qin =0.08405 m3 /s for all cyclones, air density 1.225 kg/m3 and dynamic viscosity of 178.940E-6 Pa s. The turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [74].
DPM settings
A discrete phase modeling (DPM) study has been performed by injecting 104 particles from the inlet surface with a particle density of 860 kg/m3 and with a particle size ranging from 0.025 until 5 m at a velocity equals to the gas velocity.
4.3.2
Results
76
1.25 1 0.75 0.5 0.25 0 -1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
0.25
-0.25
-1
-0.75
-0.5
-0.25
0.25
0.5
0.75
Figure 4.12: Comparison of the time averaged tangential and axial velocity between the LDA measurements, Hoekstra [74] and the current Reynolds stress turbulence model (RSM) results at 94.25 cm from the cyclone bottom.
Hoekstra [74] for more details about the used cyclone in this validation). The RSM simulation matches the experimental velocity prole with underestimation of the maximum tangential velocity, and overestimation of the axial velocity at the central region. Considering the complexity of the turbulent swirling ow in the cyclones, the agreement between the simulations and measurements is considered to be quite acceptable. The cyclone pressure drop is calculated as the pressure difference between the inlet and the average pressure across the vortex nder exit [74]. The experimental pressure drop of the cyclone can be calculated by the difference between the static pressures at the inlet and outlet [137]. A comparison of the pressure drop, the cut-off diameter (at particle density of 2740 kg/m3 ) obtained from the experimental data [74], CFD prediction is shown in Table 4.15. Table 4.15 indicates a very small deviations from the experimental values in both the calculated pressure drop and cut-off diameter. As the errors are less than 4%, so it is in the same magnitude as the experimental error [137]. The above comparison results show that the numerical model employed in this study can be used to analyze the gas ow eld and performance of the cyclone separator.
Table 4.15: Validation of the computational pressure drop and cut-off diameter Experimental [74] CFD % error Static pressure drop [N/m2 ] 300 309 3 Cut-off diameter [micron] 1 0.965 3.5
77
The dominant velocity component of the gas ow in cyclones is the tangential velocity, which results in the centrifugal force for particle separation [186]. The axial velocity is responsible more than the gravity for the transport of particles to the collection devices [29, 104]. These velocity components will be discussed in details in order to investigate the effect of the dust outlet geometry on the ow properties. Moreover, the pressure distribution in the swirling ow in these four cyclone separators will be discussed in details.
30
30
30
30
25
Tangential velocity (m/s) Tangential velocity (m/s)
25
Tangential velocity (m/s)
25
Tangential velocity (m/s)
25
20
20
20
20
15
S1 S2 S3 S4 S5 S6 S7 S8 S9
15
S1 S2 S3 S4 S5 S6 S7 S8 S9
15
S1 S2 S3 S4 S5 S6 S7 S8 S9
15
S1 S2 S3 S4 S5 S6 S7 S8 S9
10
10
10
10
0 -1
-0.5
0.5
0 -1
-0.5
0.5
0 -1
-0.5
0.5
0 -1
-0.5
0.5
79
20
15
10
S1 S2 S3 S4 S5 S6 S7 S8 S9
20
15
10
S1 S2 S3 S4 S5 S6 S7 S8 S9
20
S1 S2 S3 S4 S5 S6 S7 S8 S9
20
S1 S2 S3 S4 S5 S6 S7 S8 S9
15
15
10
10
-5 -1
-0.5
0.5
-5 -1
-0.5
0.5
-5 -1
-0.5
0.5
-5 -1
-0.5
0.5
Figure 4.13: The radial prole for the time averaged tangential and axial velocity at different sections.
1200
1000
Static pressure (N/m )
2
800
600
S1 S2 S3 S4 S5 S6 S7 S8 S9
400
200
0 0 0.5 1 -1
-0.5
-0.5
0.5
(a) Cyclone I
1400 1200
Static pressure (N/m )
-0.5
0.5
-1
-0.5
0.5
Figure 4.14: The radial proles for the time-averaged static pressure at different sections for the four cyclones.
1600 1400
Static pressure (N/m )
With dust bin Without dust bin With dipleg With dipleg & dust bin
Static pressure (N/m )
2
With dust bin Without dust bin With dipleg With dipleg & dust bin
-0.5
0.5
-0.5
0.5
1600 1400
Static pressure (N/m )
With dust bin Without dust bin With dipleg With dipleg & dust bin
-0.5
0.5
Figure 4.15: The radial prole for the time-averaged static pressure at different sections for the four cyclones.
or inverted W-shaped prole,i.e., with a maximum (V-shaped) or a dip (Wshaped) at the symmetry axis [29]. Hoekstra et al. [75] stated that the shape of the axial velocity prole is affected by the cyclone geometry. They referred the dip in the inverted W prole to the loss of swirl in the vortex nder (the friction force of the vortex nder wall attenuates the swirling ow), which results in an adverse pressure gradient at the centerline [78]. Hence, uid with fewer swirls is drawn back from the exit pipe into the cyclone. This core ow prevails throughout the entire separation space of the cyclone in spite of the attenuation of swirl in the conical part of the cyclone [78]. This explains the reason behind the inverted W-shaped prole exhibited by cyclones I (cylinder on cone), cyclone III (dipleg) and cyclone IV (dustbin plus dipleg). However, why cyclone II (with dustbin) exhibit the inverted V-shaped prole? In cyclone II, the dustbin has an equal diameter to the cyclone barrel, which means sudden expansion to the downward ow and sudden contraction to the upward ow (the ow inside the dustbin also has two streams due to the absence of a vortex stabilizer [75] which can prevent the re-entrainment of the collected particles). The upward gas ow (directed from the end of dustbin) has a higher kinetic energy especially at the cone-tip diameter (sudden contraction) which can overcome the adverse pressure gradient at the centerline (caused by the swirl attenuation in the vortex nder) and results in the inverted V axial velocity prole. One more question may appear now, why cyclones III (dipleg) and IV (dustbin plus dipleg) did not exhibit the inverted V-shaped prole? The reason can be referred to the diameter of the dustbin directly connected to the cyclone. For cyclones III and IV there is no change in the ow area at the connection, consequently no ow acceleration happens. The effect of dustbin dimensions (diameter, height) still need more investigations. The author believe, if a cone is inserted at the entrance of the cone-tip inside the dustbin of cyclone II, the axial velocity will become inverted W-shaped. However, the dimensions and location of this cone still need more investigations (cf. Obermair et al. [123] for more details).
The ow pattern
Regarding the effect of neglecting the dust outlet geometry (dustbin or dipleg or dustbin plus dipleg) in the simulating domain on the ow eld pattern, the following comments can be drawn (Fig. 4.17). 1. From the comparison between the static pressure contour plots of each cyclone and that of cyclone I, the highest value of static pressure is obtained in case of cyclone III (dipleg). The lowest value is given by cyclone II (dustbin). The highest value of the static pressure 82
35
30
Tangential velocity (m/s)
With dust bin Without dust bin With dipleg With dipleg & dust bin
Axial velocity (m/s)
25
20
With dust bin Without dust bin With dipleg With dipleg & dust bin
25
15
20
10
15
10
-5
0 -1
-0.5
0.5
-10 -1
-0.5
0.5
35
30
Tangential velocity (m/s)
With dust bin Without dust bin With dipleg With dipleg & dust bin
Axial velocity (m/s)
25
20
With dust bin Without dust bin With dipleg With dipleg & dust bin
25
15
20
10
15
10
-5
0 -1
-0.5
0.5
-10 -1
-0.5
0.5
35
30
Tangential velocity (m/s)
With dust bin Without dust bin With dipleg With dipleg & dust bin
Axial velocity (m/s)
25
20
With dust bin Without dust bin With dipleg With dipleg & dust bin
25
15
20
10
15
10
-5
0 -1
-0.5
0.5
-10 -1
-0.5
0.5
Figure 4.16: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) at the inlet region (sections S7, S8 and S9). From top to bottom: section S7-S9.
83
in cyclone IV (dustbin plus dipleg) is in between that for cyclones III and II. This indicates a slight underestimation of the pressure drop by neglecting the effect of dipleg or dustbin plus dipleg (cyclone III, IV versus cyclone I). Furthermore, a slight overestimation of the pressure drop is observed by neglecting the effect of the dustbin (cyclone II versus cyclone I). 2. The tangential velocity pattern is very similar in the four cyclones (Rankine prole). The highest value is almost the same for the four cyclones. Consequently, the collection efciency (cut-off diameter) of the four cyclones will be almost the same. Moreover, the contour plots for the tangential velocity of the four cyclones are quite similar in the main separation space (cylinder and cone). 3. The axial velocity patterns for cyclones I, III and IV have the shape of an inverted W prole while that of cyclone II has an inverted V prole, indicating different ow behavior. Close to the cone bottom there are different ow patterns as a consequence of different dust outlet geometry. Figure 4.18 shows a qualitative view of the complex ow in the four tested cyclones with the streamtraces plots of the time averaged velocities colored by the time-average axial velocity. The swirling, downward ow at the outer region of the cyclone is clearly visible. Near the bottom of the cyclone, it is diverted into an upward ow near the cyclone center. In cyclone II (with dustbin) and IV (with dustbin and dipleg), the ow behavior in the dustbin is quite different because of the different length of the dustbin.
The performance
Figure 4.19 presents the grade efciency curves (GECs) for the four tested cyclones. As expected, the frictional efciencies of all the cyclones are seen to increase with the increase in particle size. The shapes of the grade collection efciency curves of all models have a so-called S shape [195]. It is clear from Fig. 4.19 that the effect of the cyclone dustbin or dipleg, on the cut-off diameter (particle diameter of 50% collection efciency) is small. Neglecting the effect of dust outlet geometry in the cyclone simulation slightly overestimates the cut-off diameter (when compare it with cyclone I). A deeper look to the GECs indicates the variation of the collection efciency for particles diameters larger than 1.5 m, with a higher efciency for cyclone II and the lower efciency for cyclone I. This behavior is due to the increase in the separation space and the change in the maximum tangential velocity. For particles with diameters less than 0.8 m, 84
Figure 4.17: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From left to right: the static pressure (N/m2 ), the tangential and the axial velocity (m/s).
85
Figure 4.18: The streamtraces plots for the time averaged ow variables, colored by the average axial velocity (m/s).
86
1 Without dustbin With dustbin With dipleg With dustbin plus dipleg 0.8
Collection efficiency
0.6
0.4
0.2
4 5
Figure 4.19: The grade efciency curves for the four cyclones.
4.3.3 Conclusions
Four cyclones have been simulated using the Reynolds stress model (RSM), to study the effect of the dust outlet geometry on the cyclone separator performance, ow pattern and velocity proles. The following conclusions have been obtained. 87
4.4 Closure
From the previous sections, it becomes clear that the most signicant factors are four (the vortex nder diameter, the inlet width, the inlet height and the total cyclone height (cone height)). Consequently, any minor modication in these particular factors will result in a considerable change in the cyclone performance. The effect of the cone-tip diameter on the cyclone performance is insignicant. Since the main target of this thesis is to optimize the cyclone performance, it is accepted to exclude the dust hopper (dustbin or dipleg) from the simulation domain and use the obtained CFD data to obtain the optimized cyclone design.
88
investigations on the particle collection efciency of small cyclones operating at high-ow rates. Special emphasis was given to the effects of the barrel height and the vortex nder length on the particle collection efciency. The length ratios of the barrel height h and vortex nder length S to the cyclone body diameter D were varied from 0.75 to 4.5 and from 0.5 to 1.5, respectively. Pressure drop decreased substantially either as the barrel height h became longer or as the vortex nder length S became shorter. It was also found that the difference between the cyclone barrel height h and the vortex nder length S affects the particle collection characteristics signicantly. The optimum performance will be obtained if (h S )/D = 1.0.
The performance of a cyclone, with different vortex nders, was evaluated by Lim et al. [102] to examine the effect of the vortex nder shape on the characteristics of the collection efciency. Four cylinder-shaped and six cone-shaped vortex nders were designed and employed to compare the collection efciencies of the cyclone, at ow rates of 30 and 50 l/min. The cylinder-shaped-vortex nders had different diameters and the coneshaped vortex nders had different cone lengths. The result indicates that two cone-shaped vortex nders, with different diameters, had the collection efciencies between those of the cylinder-shaped vortex nders with the same diameter, and that a smaller pressure drop per ow rate unit could be achieved for the cone-shaped design, but the cone length did not affect the collection efciency and pressure drop of the cyclone. Raou et al. [140] duplicated numerically the same study of Lim et al. [102] with limited details about the effect of the gas outlet diameter on the ow eld 90
5.1. Introduction
pattern and velocity prole. You-hai et al. [188] simulated the three-dimensional gas-phase ow eld in the cyclone separator with different vortex nder diameters. The results show that when the diameter of the vortex nder decreases, the downward ow decreases and the tangential velocities of the whole cyclone separator increase, but at the cost of the pressure drop increase. Horvath et al. [79] refereed the reason behind the two classes of axial velocity in cyclone separator (class V and class W) to the inuence of the vortex nder diameter. The pressure drop and collection efciency of a swirl tube with different vortex nder geometries were studied numerically by Jian and You-hai [87]. The gas ow elds were simulated by the Reynolds stress model (RSM) and the stochastic tracking approach in discrete phase model (DPM). The results indicate that the decrease of the vortex nder diameter leads to higher tangential velocity, which helps to improve the separation efciency. The back ow can be observed in converging coned-shaped vortex nder, meanwhile the diverging coned-shape vortex nder can make the ow move smoothly to reduce pressure drop. Ficici et al. [57] performed an experimental study using three cylindershaped vortex nders with diameters of 80, 120 and 160 mm. They investigated the effects of gas inlet velocity, the vortex nder diameter Dx and length S on the cyclone performance. They reported a linear relationship between the length of the vortex nder and the pressure loss. Khalkhalia and Sakhania [94] performed a multi-objective optimization of a cyclone vortex nder shape using CFD simulations data set. Two meta-models based on the evolved group method of data handling (GMDH) type neural networks are used as tness functions for Pareto-based optimization. In summary, all previous studies reported the signicant effect of the vortex nder dimensions on the cyclone performance and ow pattern. Nevertheless, the previous studies are not coherent, and did not present sufcient details about the effect of these two geometrical parameters on the pressure drop and the cut-off diameter. Moreover, detailed studies about the effect of the vortex nder dimensions on the velocity proles are scarce in the literature. The present study is intended to computationally investigate the effect of increasing the vortex nder diameter Dx and length S on the pressure drop and cut-off diameter and to obtain more details about the ow eld pattern and velocity proles using the large eddy simulation (LES) methodology. 91
The numerical simulations were performed on ve cyclones with different vortex nder diameters (at constant S ) and ve cyclones with different vortex nder length (at constant Dx ). Figure 5.1 and Table 5.1 give the cyclones geometrical dimensions.
5.2.2
Solver settings
The simulations started with unsteady simulation using the Reynolds stress turbulence model with a time step of 1E-4 for initialization of the ow eld. Afterwards, the turbulence model switched to the large eddy simulation with a time step of 1E-5s using implicit coupled solution algorithm. The selected time step results in an average inlet Courant number of 0.0288 for the tested cyclones. The cyclones volumes and the corresponding residence times for the tested cyclones are given in Table 5.2. The minimum value of residence time is 0.0812s i.e., the used time step is just a small fraction of the average residence time. This conrms that the used time step can reveal the transient phenomena properly. For the tested cyclones the maximum values of the courant number are less than 0.2. This veries again that the choice of the time step was proper (cf. Sec. 4.2.1.3, page 56 for more details).
Table 5.1: The geometrical dimensions of the tested cyclonesa
Dimension Inlet height, a Inlet width, b Barrel height, h Total cyclone height, Ht Cone tip-diameter, Bc Vortex nder diameter, Dx S/D = 0.5 Cycloneb dimension/D 0.375 0.2625 1.5c 4.0 0.375 0.30 0.35 0.40 0.45 0.50 0.5 0.625 0.875 1.0 1.0 (h S )/D Dx /S
Dx /D = 0.5
D1 D2 D3 D4 D5 S1 S2 S3 S4 S5
1.0 1.0 1.0 1.0 1.0 1.0 0.875 0.625 0.5 1.0c
0.6 0.7 0.8 0.9 1.0 1.0 0.8 0.625 0.5 1.0
a Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at a distance Li = D from the cyclone center, cf. Fig. 5.1(a). b c
Cyclone D5 and S1 are identical. cyclone S6 has different barrel height h/D = 2.0 .
92
a The total number of hexahedral cells after the grid independency study, cf. Sec. 5.2.4 b The residence time t res = V/Qin where V is the cyclone volume and Qin is the gas ow rate (50 l/min).
a 2 Euler number is the dimensionless pressure drop Eu = P/( 1 2 Vin ) where P is the static pressure drop, is the gas density, Vin is the gas inlet velocity. b The percentage difference between the coarsest and nest grid. c The percentage difference between the ne and nest grid.
93
puted using three levels of grids in order to estimate accurately the order of convergence and check that the solution is within the asymptotic range of convergence [160]. The GCI is based upon a grid renement error estimator derived from the theory of the generalized Richardson extrapolation [55, 160]. The GCI is a measure of how far the computed value is away from the value of the asymptotic numerical value. Consequently, it indicates how much the solution would change with a further renement of the grid. A small value of GCI indicates that the computation is within the asymptotic range (cf. Sec. 4.2.1.3 for more details). Table 5.4 presents the grid convergency calculations using GCI method for three grid levels for cyclone S1 as an example for the tested cyclones. The following conclusions have been obtained from the GCI analysis [55]: The results are in the asymptotic range, because the obtained values for are close to unity. The ratio R is less than unity this means monotonic convergence [2]. There is a reduction in the GCI value for the successive grid renene ments (GCIne 12 < GCI23 ) for the two variables (Eu and x50 ). This indicates that the dependency of the numerical results on the cell size has been reduced and a grid independent solution has been achieved. Further renement of the grid will not give much change in the simulation results. For the two variables (Eu and x50 ), the extrapolated value is only slightly lower than the nest grid solution. Therefore, the solution has converged with the renement from the coarser grid to the ner grid [2]. Figure 5.2 presents a qualitative proof that the obtained results are in the asymptotic range.
Table 5.4: Grid convergency calculations using GCI method and three grid levels for cyclone S1
i 0c 1 Eu 2 3 0 1 x50 2 3 Ni 1793459 1174029 714029 1793459 1174029 714029 fi 3.1336 3.1450 3.1800 1.1803 3.3750 1.2703 1.3180 1.1517 1.4360 1.1803 1.9850 0.5490 0.3823 14.4222 0.1180 0.0895 4.5223 0.2342 1.0895 0.1950 0.0613 1.8228 ri,i+1 ei,i+1 i,i+1 GCIne i,i+1 % Ra b
1.1517
0.0350
0.0111
a R= / . 12 23 b = rp GCI 12 /GCI23 . 12 c The value at zero grid space (h 0). i=1, 2 and 3 denote the calculations at the ne, medium and coarse mesh respectively.
94
2.2
2 3.3
Euler number
3.1
1.4
(h --> 0)
N -1
Figure 5.2: Qualitative representation of the grid independency study. The Euler number and the cut-off diameter for cyclone S1 at the three grid levels. N 1 is the reciprocal of the number of cells, h 0 means the value at zero grid size (cf. Table 5.4). To obtain a smooth curve; the spline curve tting has been applied in Tecplot post-processing software. Table 5.5: The position of different plotting sections Section z`/Da
a
S1 2.75
S2 2.5
S3 2.25
S4 2
S5 1.75
S6 1.5
Cut-off diameter
S7 1.25
S8 1.0
S9 0.75
In the outer region, due to the sharp drop in the velocity magnitude in the near wall region, the distribution is different but the maximum tangential velocity is similar at all sections for the same cyclone (S1-S5). The maximum tangential velocity increases with decreasing the vortex nder diameter. The cyclones S1 to S5 show the inverted W prole but cyclones D1 to D3 show the inverted V prole. The radial proles of the time averaged static pressure are given in Fig. 5.3. Like for the tangential velocity, the axial variations of static pressure are very small for the same cyclone. The maximum value of the static pressure decreases when the vortex nder diameter is increased for cyclones D1 to D5. However, the maximum value of the static pressure slightly increases when the vortex nder length is decreased for cyclones S1 to S4. Cyclone S5 (Dx /D = 0.5, S/D = 0.5, h/D = 2.0) differs than cyclone S4 in only the barrel height (h/D = 1.5). From Fig. 5.4, a slight difference in the maximum tangential velocity between the two cyclones, which predict close values for the collection efciency. Moreover, the axial velocity proles for the two cyclones are close but less central dip in the axial velocity is exhibited in cyclone S5. The distinct difference between cyclones S4 and S5 is that in the maximum static pressure, where cyclone S5 indicates a reduction in the maximum static pressure. This indicates that the effect of changing the barrel height is more signicant than that of the vortex nder length. The sharp changes in the radial proles appear in Fig. 5.4 are due to crossing the vortex nder.
5.3.2
The ow pattern
Figure 5.5 shows the contour plots of the time-averaged static pressure, tangential and axial velocity for cyclones D1-D5. The time-averaged static pressure decreases radially from the wall to the center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, whereas the gradient in axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the low-pressure zone at the cyclone center (twisted cylinder). Two vortical motions are exist one moving down (outer vortex) and the other moving up (inner vortex). The highest value of the static pressure decreases with increasing the vortex nder diameter. Consequently, a smaller pressure drop can be expected when increasing the vortex nder diameter. The tangential velocity pattern is very similar in all cyclones (Rankine prole). The highest value decreases with increasing the vortex nder diameter. Consequently, a better collection efciency can be expected when decreasing the vortex nder 96
D1
16 16 12 12
D2
16 12
D3
14 12 10
D4
12 10
D5
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-4 -0.015
-0.015
-4 -0.015
0 -0.015
0 -0.015
15
12
10
10
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
7.5
S1 S2 S3 S4 S5 S6 S7 S8 S9
2.5
0 -2 -2 -2.5
97
-5 -0.015
-0.01
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
-5 -0.015
-0.01
0.01
0.015
500
400 320
280
200
160
300
200
160
120
50
200
100
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
100
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
80
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
80
40
-40 -0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
0 -0.015
-0.015
-0.015
-100 -0.015
Figure 5.3: The radial prole for the time-averaged tangential and axial velocity and static pressure at different sections for cyclones D1 - D5.
S1
12 12 10 10
S2
12 10
S3
12 10
S4
12 10
S5
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
0 -0.015
0 -0.015
0 -0.015
0 -0.015
0 -0.015
10
10
10
10
10
7.5
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
2.5
-2 -2.5 -4 -5 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015 -5 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015 -5 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-2
100
100
100
100
98
200
200
200
200
200
150
150
150
150
150
100
50
50
50
50
50
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-100 -0.015
-100 -0.015
-100 -0.015
-100 -0.015
-100 -0.015
Figure 5.4: The radial prole for the time-averaged tangential and axial velocity and static pressure at different sections for cyclones S1 - S5.
diameter. Figure 5.6 shows the contour plots of the time-averaged static pressure, tangential and axial velocity for cyclones S1-S5. The highest value of the static pressure slightly decreases with increasing the vortex nder length. The tangential velocity pattern is very similar in all cyclones (Rankine prole). The highest value is very closed for cyclones S1 to S4 but there is a small difference between cyclones S4 and S5. The axial velocity patterns for the ve cyclones (S1 to S5) have the shape of an inverted W prole. Cyclone S5 exhibit a less central dip in the axial velocity.
D1
D2
D3
D4
D5
100
Figure 5.5: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section for cyclones D1 - D5. From top to bottom: the static pressure (N/m2 ), the tangential velocity (m/s) and the axial velocity (m/s). From left to right cyclone D1 - D5.
S1
S2
S3
S4
S5
101
Figure 5.6: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section for cyclones S1 - S5. From top to bottom: the static pressure (N/m2 ), the tangential velocity (m/s) and the axial velocity (m/s). From left to right cyclone S1 - S5.
on the axial velocity prole is very small in comparison with the vortex nder diameter. The shape of the axial velocity prole is affected by the cyclone geometry [75]. Hoekstra et al. [75] referred the dip in the inverted W prole to the loss of swirl in the vortex nder (the friction force of the vortex nder wall attenuates the swirling ow), which results in an adverse pressure gradient at the centerline [78]. Hence, uid with fewer swirls is drawn back from the exit pipe into the cyclone. This core ow prevails throughout the entire separation space of the cyclone in spite of the attenuation of swirl in the conical part of the cyclone [78]. This explains the reason behind the inverted W-shaped prole exhibited by cyclones S1 to S5. However, why decreasing the vortex nder diameter gradually change the axial velocity prole from inverted W to exhibit the inverted V-shaped prole? The answer is, when the vortex nder diameter decreases, the swirl in the cyclone increases (as is clear from the tangential velocity proles for cyclones D1 to D5). Consequently, the ow can overcome the adverse pressure gradient and exhibit the inverted V prole for cyclone D1.
5.3.4
In order to estimate the effect of the vortex nder dimensions on the performance parameters, the Euler number (the dimensionless pressure drop) have been calculated. A discrete phase modeling (DPM) study has been performed by injecting 104 particles from the inlet surface with a particle density of 860 kg/m3 and with a particle size ranging from 0.025 until 5 micron. Figure 5.8 shows a sharp decrease of the Euler number (dimensionless pressure drop) with increasing the vortex nder diameter Dx and a small increase with increasing the vortex nder length S . This behavior can be explained as follows. The pressure drop in the cyclone is composed of three main contributions [52]: (1) the pressure drop at the inlet section. (2) the pressure drop in the cyclone body due to swirling motion and due to wall friction, this contribution may increase with increasing the cyclone barrel height (cyclone S5) or with increasing the vortex nder length (cyclones S1 - S4) as the wall friction will increase due to friction with a larger wall surface, or decreases as the vortex strength will decrease because the maximum tangential velocity decreases. (3) the main contribution to the cyclone pressure drop is the energy loss in the vortex nder, which mainly depends on the maximum tangential velocity in the cyclone. As is clear from Fig. 5.7 the maximum tangential velocity decreases with increasing the vortex nder diameter. As the inlet section is the same in the all tested 102
500
200
400
150
300
100
200
50
100
-50
-0.015
-100 -0.015
12 14 10 12 8
10
-2 -0.015
-2 -0.015
2.5
-2.5 0
-2 -0.015
-0.01
0.01
0.015
-5 -0.015
Figure 5.7: Comparison between the radial proles for the time averaged static pressure, tangential and axial velocity at section S6.
103
12 11 10 9
0.8
Euler number (Dx) Stk50 x 10 3 (Dx) Euler number (S) Stk50 x 10 3 (S)
2.5
Euler number
8 7 6 5
1.5
0.5 4 3 0
0.3
0.35
0.4
0.45
0.5
Dx
Figure 5.8: The variation of the Euler number and the Stokes number with the vortex nder dimensions.
The trend of increasing the Stokes number with increasing the vortex nder diameter is quite reasonable, as the centrifugal force affecting particles attenuates when the swirl intensity (maximum tangential velocity) decreases (Fig. 5.8). The Stokes number slightly increases as the vortex nder length is increased (cyclone S1 - S4). The insignicant change of Stokes number is quite reasonable because of very limited changes in the ow pattern (cf. Fig. 5.7). The small increase in the Stokes number can be explained with the aid of the contours plots given in Fig. 5.6. The zone of peak axial velocity increases with increasing S . Consequently, the possibility of carrying bigger particles to escape with the upward ow slightly increased. Cyclone S5 depicts a sudden reduction in the Stokes number 104
Stk50 x 10
values in comparison with cyclone S4. This drop is due to separation space increase in cyclone S5, and the possibility of particles to be captured increases. Although, the vortex strength decreased with a small amount due to the reduction in the maximum tangential velocity. Figure 5.9 presents the variation of the cyclone performance parameters with the ratio Dx /S . It is clear that the effect of changing Dx /S depends on the variables (Dx or S ). Consequently, there are two curves per performance parameter (Euler number and Stokes number). The common point in Fig. 5.9 is cyclone D5 (=S1).
12 11 10 9
Euler number (Dx) Stk50 x 10 3 (Dx) Euler number (S) Stk50 x 10 3 (S)
Euler number
8 7 6 5 4 3
0.6
0.8
D x /S
Figure 5.9: The variation of the Euler number and the Stokes number with the x ratio of D (cyclone S5 is excluded). S
The variation of the performance parameters with the ratio of (h S )/D is presented in Fig. 5.10 for cyclones S1 - S5. Increasing (h S )/D decreases both the Euler number and Stokes number. However, the values of the performance parameters depend on the value of the barrel height as is clear from the values for cyclone S5. In order to obtain the Euler number-Stokes number relationship, Fig. 5.11 has been drawn. It indicates a general relationship (trend) between the two dimensionless numbers irrespective to the geometrical parameters values. Two second-order polynomials have been proposed by Elsayed and Lacor [54, 56], Eqs. 5.1 and 5.2. Equation 5.2 presents a good matching for the performance parameters 105
Stk50 x 10
3.8
Euler number (S1-S4) Stk50 x 10 3 (S1-S4) Euler number (S5) 3 Stk50 x 10 (S5)
2.4
2.3
Euler number
2.2
0.5
0.625
0.75
0.875
(h-S)/ D
Figure 5.10: The variation of the Euler number and the Stokes number with the S . ratio of hD
for cyclones D1 -D5. For smaller values of Euler number, there is underestimation of the Stokes number if the Eq. 5.2 is applied. The difference between the two correlations can be referred to two reasons. Firstly, Eq. 5.1 is based on the CFD simulations data for both Euler number and Stokes number, whereas Elsayed and Lacor correlation [56] (Eq. 5.2) was obtained from experimental data for Euler number values and Iozia and Leith model for the Stokes number values. Secondly, Eq. 5.1 is limited to only four geometrical parameters. The other three factors are xed, h = 1.5, S = 0.5 and Bc = 0.375. Consequently, Eq. 5.1 is not suitable to t the Euler number - Stokes number relationship for cyclones S1 - S5 because the values of S is away from the range of applicability. Stk50 = 100.3533(log10 (Eu)) Stk50 = 100.3016(log10 (Eu))
2
1.1645log10 (Eu)2.3198
Stk50 x 10
3.6
(5.1) (5.2)
0.9479log10 (Eu)2.5154
5.4 Conclusions
Nine cyclones of different vortex nder dimensions (diameter and length) have been simulated using the large eddy simulation (LES) methodology, 106
5.4. Conclusions
3
Simulation Correlation 1 Correlation 2
2.5
Stokes number x 10 3
1.5
Euler number
Figure 5.11: The variation of the Stokes number with the Euler number for cyclones D1D5.
to study the effect of the vortex nder dimensions on the performance and ow pattern. The following conclusions have been obtained. The maximum tangential velocity in the cyclone decreases with increasing the vortex nder diameter. A negligible change is noticed with increasing the vortex nder length. Very limited axial variations in the ow variables are reported with changing the vortex nder dimensions for the same cyclone. Increasing the vortex nder length, makes a small change in both the static pressure, axial and tangential velocity proles. However, decreasing the vortex nder diameter gradually changes the axial velocity prole from the inverted W to the inverted V prole. Decreasing the cyclone vortex nder diameter, increases the maximum tangential velocity. The maximum tangential velocity approaches asymptotically 1.589 times the inlet velocity when decreasing the vortex nder diameter. The Euler number (dimensionless pressure drop) decreases with increasing the vortex nder diameter Dx . Increasing the vortex nder length S slightly increases the Euler number. The Stokes number increases with increasing the vortex nder diameter, because the centrifugal force affecting particles attenuates when the swirl intensity (maximum tangential velocity) decreases. The Stokes number slightly increases as the vortex nder length is 107
108
Eu = 11.3
+ 2.33
(6.1)
implying proportionality with the square of the inlet area. Ramachandran et al. [139] on the other hand proposed, Eu = 20 ab 2 Dx
S D H h Bc DD D 1/3
(6.2)
i.e. a linear relation with the inlet area. Iozia and Leith [84, 85] presented a correlation to estimate the cut-off diameter d50 and found proportionality to (a b)0.61 . The importance of inlet dimensions becomes clearer after the study of natural length (or vortex length) by several researchers, e.g., Alexander [1]. The cyclone has two spiral motions, outer and inner. In the reverse ow cyclone, the outer vortex weakens and changes its direction at a certain axial distance Ln from the vortex nder [29]. This distance is usually called the turning length, natural length or vortex length of the cyclone. The inlet area is one of the relevant parameters inuencing the natural length. Alexander [1] found that Ln decreased proportionally to the inlet area (Ai = a b) but the opposite trend has been also reported [29]. 109
The effects of cyclone inlet on the ow eld and performance of cyclone separators have been numerically investigated by Zhao et al. [198]. They compared the performance of two types of cyclones with the conventional single inlet and spiral double inlets using the Reynolds stress turbulence model. The results show that the new type cyclone separator with spiral double inlet can improve the symmetry of gas ow pattern and enhance the particle separation efciency. While their nding is for double inlets cyclone, it supports the importance of the effect of the inlet section dimensions on the performance of cyclone separator. The signicant effects of the cyclone inlet dimensions on the cyclone performance have been acknowledged in many articles [e.g., 5]. For two inlets cyclone separators, Zhao [193] reported the possibility of increasing the cyclone efciency without signicantly increasing the pressure drop by improving the inlet geometry of the cyclone. The effect of inlet section angle has been tested by many researchers. Qian and Zhang [137] computationally investigated the effect of the inlet section angle. The pressure drop of the cyclone decreases to a 30% lower value than that for conventional cyclone, if the inlet section angle becomes 45 . However, Qian and Wu [136] reported only 15% reduction in the pressure drop for = 45 .
In summary, all articles mentioned above did not study the effect of the inlet height or width dimensions on the performance and ow pattern but they studied the effect of the inlet congurations (inclined instead of tangential), or the effect of the number of inlets (single or double) or the shape of the inlet section (rectangular duct or nozzle). The present study is intended to computationally investigate the effect of increasing the cyclone inlet width and height on the pressure drop and cut-off diameter and obtaining more details about the ow eld pattern and velocity proles. The study will be done using RANS and the Reynolds stress turbulence model (RSM) will be used as a closure turbulence model. 110
where Dx is the vortex nder diameter, D is the cyclone body diameter, and Ain is the inlet cross-sectional area. For industrial cyclones, the geometrical swirl number usually varies between 1 and 5 [74]. Table 6.2 shows that, the tested cyclones cover this range. Implying that the obtained results can be applied to the industrial cyclones.
A1 A2 A3 B1 B2 B3
Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at a distance Li = D from the cyclone center.
an acceptable value for the current simulation for accurate results and achieve scaled residuals less than 1e-5 for all variables. The selected time step results in an average inlet Courant number of 29.25, 28.88, 21.67, 30.40, 21.45 for cyclones A1, A2, A3, B1 and B3 respectively. However, as the solver is a segregated implicit solver, there is no limitations on the Courant number for stability.
6.2.4
CFD grid
Figure 6.1(b) shows the surface grid of cyclone A2 used in this study. The hexahedral computational grids were generated using GAMBIT grid generator. A grid independency study has been performed for the ve tested cyclones. Three grid levels for each cyclone have been tested, to be sure that the obtained results are grid independent. For example, for cyclone A2 (B2) three levels of meshes with respectively 490164, 714029 and 1174029 cells have been used. The computational results are presented in Table 6.3. As seen, the maximum difference between the results is less than 5%, so the grid with 490164 cells can be considered as adequate [146]. It has been obTable 6.2: The details of the ve tested cyclones Cyclone Number of cells Sg Cyclone volume x105 [m3 ] tres [s] Inlet velocity [m/s]
The total number of hexahedral cells after the grid independence study
112
6.3. Results
Dx
Le
h Li D b
Ht
Bc
Figure 6.1: Schematic diagram and surface mesh for the cyclone separator
served that even 490164 grid provides a sufcient grid independency. However, for excluding any uncertainty, computations have been performed using the 714029 cells grid, where the total number of grid points was not that critical with respect to the computation overhead [10].
Table 6.3: The details of the grid independence study for cyclone A2
Number of cells 490164 714029 1174029 % difference
6.3 Results
6.3.1 The axial variation of the ow properties
The tangential velocity is the dominant component of the gas ow in cyclones, which results in the centrifugal force for particle separation [186]. 113
Moreover, the development of axial velocity prole in axial direction will be analyzed for the ve cyclones. Nine sections are used to plot the velocity proles as shown in Table 6.4. Figures 6.2, 6.3 and 6.4 present the radial proles of the time-averaged static pressure, tangential and axial velocity at 9 axial stations. As expected, the tangential velocity proles exhibit the so-called Rankine vortex, which consists of two parts, an outer free vortex and an inner solid rotation in the center (Fig. 6.3). The tangential velocity distribution in the inner region is rather similar at different sections for the same cyclone. In the outer region, due to the sharp drop in velocity magnitude in the near wall region, the distribution is different and the change in the value of maximum tangential velocity is rather limited. Generally, the tangential velocity distribution varies only slightly with axial positions for the same cyclone, which is also reported in other articles [e.g., 66, 127, 163, 186]. This means that, if the tangential velocity increases at one section of the cyclone, it will increase at all other sections. The same conclusion can be drawn from the radial prole of static pressure with higher values of pressure drop expected for cyclone A1 in comparison with cyclones A2 and A3. Cyclone B1 also depicts higher values of pressure in comparison with cyclones B2 and B3 (Fig. 6.2). The axial velocity prole has the shape of an inverted W for all cyclones except B3 with the shape of an inverted V, as a result of the change in the ow eld pattern due to very wide inlet section.
Table 6.4: The position of different plotting sections Section z`/D
S1 2.75
S2 2.5
S3 2.25
S4 2
S5 1.75
S6 1.5
S7 1.25
S8 1.0
S9 0.75
To evaluate the effect of increasing the cyclone inlet width on the tangential and axial velocity proles, the tangential and axial velocity proles at section S9 (close to the inlet section) for the three cyclones (both for A cyclones and B cyclones) are compared in Fig. 6.5. As is clear from Fig. 6.5 the variation of axial velocity close to the wall is limited when changing the inlet width or height. The axial velocity proles for the three cyclones are very similar except at the central region. The most important is the effect of cyclone inlet height or width on the tangential velocity (proportional to the centrifugal force, which is the main force in the separation process). Increasing the cyclone inlet width or height decreases the maximum tangential velocity. Cyclone A1 and B1 have the maximum tangential velocity in comparison with other cyclones. This means that decreasing the cyclone inlet dimension will enhance the collection efciency. 114
6.3. Results
B1
250
200
Static pressure [N/m ]
2
150
100
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50 -1
-0.5
0
Radial position [-]
0.5
-0.5
0
Radial position [-]
0.5
A2
200 200
B2
150
Static pressure [N/m ] Static pressure [N/m ]
150
100
100
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50 -1
-0.5
0
Radial position [-]
0.5
-50 -1
-0.5
0
Radial position [-]
0.5
A3
200 125
B3
150
Static pressure [N/m ] Static pressure [N/m ]
100
75
100
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
25
-50 -1
-0.5
0
Radial position [-]
0.5
-25 -1
-0.5
0
Radial position [-]
0.5
Figure 6.2: The radial prole for the time-averaged static pressure at different sections. Note: A2=B2.
116
6.3. Results
A1 B1
A2
B2
A3
B3
Figure 6.3: The radial prole for the time-averaged tangential velocity at different sections. Note: A2=B2.
117
A2
B2
A3
B3
Figure 6.4: The radial prole for the time-averaged axial velocity at different sections. Note: A2=B2.
118
6.3. Results
300
250
200
Static pressure [N/m ]
2
150
100
50
-50 -1
-0.5
0
Radial position [-]
0.5
-0.5
0
Radial position [-]
0.5
Figure 6.5: Comparison between the radial proles for the time averaged static pressure, tangential and axial velocity at section S9.
119
at this region. The author believes that the extent of asymmetry is not just a function of the cyclone geometry and boundary condition at the cone bottom, but also the operating condition like the gas ow rate. This asymmetry causes the obtained axial velocity variations (spot-like distribution close to the cone bottom). (3) Another reason can be the boundary condition at the gas outlet (vortex nder exit). A fully developed ow (outow boundary condition) is assumed. Actually, there is no guarantee that the ow is fully developed. May be the ow is still developing throughout a long distance after the vortex nder exit. However, Wang et al. [178] reported that the effect of the gas outlet length on the simulated velocity becomes insignicant after a distance of 1/2 the cyclone diameter. In spite of the above-mentioned interpretations, the following parameters need to be investigated to accurately explain the reason of this ow behavior: (i) The effect of including the dustbin in these simulations. (ii) The effect of the gas outlet tube length. (iii) The effect of the operating condition; e.g., the gas Reynolds number (it includes the effect of inlet velocity, gas density, and viscosity), the operating pressure, and temperature. Moreover, it may be worthwhile to perform these investigations experimentally (using the particle image velocimetry (PIV) technique or any other ow visualizations technique) to check also if these phenomena (observations) come only from the numerical solution (schemes, settings, boundary conditions, etc.).
6.3.3
In order to estimate the effect of cyclone inlet dimensions on the performance parameters, the pressure drops have been calculated and presented in Fig. 6.8. A discrete phase modeling (DPM) study has been performed by injecting 104 particles from the inlet surface with a particle density of 860 kg/m3 and with a particle size ranging from 0.025 until 5 micron. 6.3.3.1 The effect of the inlet height Figure 6.8 shows a rapid decrease in the pressure drop when increasing the inlet height for 0.25 a/D 0.4 and a smaller decrease for a/D 0.4. This behavior can be explained as follows. The pressure drop in the cyclone is composed of three main contributions: (1) the pressure drop at the inlet section (decreased by increasing the inlet dimensions). (2) the pressure drop in the cyclone body due to swirling motion and due to wall friction, this contribution decreases with increasing the cyclone inlet height (as the vortex strength will decrease). (3) the main contribution to the cyclone 120
6.3. Results
A1
The static pressure N/m2
A2
A3
Figure 6.6: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From top to bottom : the static pressure N/m2 , the tangential velocity m/s and the axial velocity m/s. From left to right cyclone A1 through cyclone A3.
121
B1
The static pressure N/m2
B2
B3
Figure 6.7: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From top to bottom : the static pressure N/m2 , the tangential velocity m/s and the axial velocity m/s. From left to right cyclone B1 through cyclone B3.
122
6.3. Results
pressure drop is the energy loss in the exit tube, which mainly depends on the maximum tangential velocity in the cyclone. As is clear from Fig. 6.5 the maximum tangential velocity decreases with increasing cyclone inlet height. Consequently, in general the total cyclone pressure drop will decrease with increasing the cyclone inlet height. Figure 6.8 also depicts the pressure drop for the three cyclones using four different mathematical models; Muschelknautz method of modeling (MM) [29, 175], Casal and Martinez-Benet [21], Shepherd and Lapple [157], Stairmand [166] indicating the same overall trend obtained using RSM simulations for the pressure drop with better matching between the MM model and the current RSM simulation. The effect of the cyclone inlet height on the cut-off diameter (particle diameter of 50% collection efciency) is shown in Fig. 6.8. The general trend is an increase of the cut-off diameter with increasing cyclone inlet height, due to weakness of the vortex. Figure 6.8 also depicts the cut-off diameter for the three cyclones using two mathematical models; Iozia and Leith [85] and Rietema [142] indicating the same overall trend obtained using RSM simulations for the cut-off diameter with exact matching between the current RSM results and Rietema model.
6.3.3.2 The effect of the inlet width Figure 6.8 shows a rapid decrease in the pressure drop when increasing the inlet width for 0.15 b/D 0.27 and a smaller decrease for b/D 0.27. Generally, both the cyclone pressure drop and the cut-off diameter decrease with increasing the cyclone inlet width. The used mathematical models again indicate the same overall trend obtained using RSM simulations for cut-off diameter and pressure drop. Figure 6.8 indicates the need of applying a multi-objective optimization procedure to get the optimum value for the inlet dimensions. Both the collection efciency (cut-off diameter) and the pressure drop in cyclone separator are important objective functions to be optimized simultaneously [146]. The effects of changing the cyclone inlet dimensions on pressure drop and collection efciency are opposite. Increasing the inlet width will save more driving power but leads to reduced collection efciency (larger cut-off diameter). From the graph, the optimum value will be close to b/D = 0.25 with large reduction in pressure drop (energy losses) and small increase in cut-off diameter. Elsayed and Lacor [50] estimated the optimum value for inlet width (b/D) equals 0.236. 123
6.3.3.3 The inlet height versus the inlet width Figure 6.9 shows that, the impact of changing either the inlet height or width on the pressure drop is almost the same. However, the effect of changing the inlet width on the cut-off diameter is more signicant in comparison with that of the inlet height. When plotting the pressure drop against the ratio of inlet width to inlet height, it becomes clear the optimum range of b/a is from 0.5 until 0.7.
6.4 Conclusions
Five cyclones of different inlet width and height have been simulated, using the Reynolds stress model (RSM), to study the effect of cyclone inlet dimensions on the cyclone separator performance and ow pattern. The following conclusions have been obtained. The maximum tangential velocity in the cyclone decreases with increasing both the cyclone inlet width and height. No acceleration occurs in the cyclone space (the maximum tangential velocity nearly constant throughout the cyclone). The axial variation of both the static pressure and axial velocity is very limited. Increasing the cyclone inlet width or height decreases the pressure drop at the cost of increasing the cut-off diameter. So an optimization procedure is needed to estimate the optimum value of inlet dimensions. Wider inlet cyclones (b/D > gap between the cyclone barrel and the vortex nder are not preferred. The effect of changing the inlet width on the cut-off diameter is more signicant in comparison with that of the inlet height. The optimum ratio of the inlet width to the inlet height b/a is from 0.5 to 0.7.
124
6.4. Conclusions
1000 900
RSM MM Casal Shepherd Stairmand RSM Iozia Rietma
2
Pressure drop Cut-off diameter
0.25
0.3
0.35
0.4
0.45
0.5
a/D
1000 900
RSM MM Casal Shepherd Stairmand RSM Iozia Rietma Pressure drop Cut-off diameter
0.15
0.2
0.25
0.3
0.35
b/D
Figure 6.8: The effect of inlet dimensions on the pressure drop and cut-off diameter using CFD simulations and different mathematical models.
125
0.2
0.25
0.3
0.35
0.4 2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.55
0.25
0.3
0.35
0.4
0.45
0.5
400
Pressure drop [N/m ] Cut-off diameter [micron]
2
300
2 1.8 1.6
200 1.4 1.2 100 1 0.8 0.6 0 0.2 0.4 0.6 0.8 1 0.4 1.2
b/a
Figure 6.9: Comparison between the effect of inlet height and width on the pressure drop and cut-off diameter using CFD simulations.
126
The numerical simulations were performed on four cyclones with different barrel heights (at constant cone height) and three cyclones with different cone heights (at constant barrel height). Figure 7.1 and Table 7.1 give the cyclones dimensions.
7.2.2
Boundary conditions
Table 7.2 shows the details of the boundary conditions. The air volume ow rate Qin =50 L/min for all cyclones, air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s. The turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [52, 74]. 128
C1 C2 C3 B1 B2 B3 B4
Barrel height, h
hc /D = 2.5
a Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at a distance Li = D from the cyclone center. b The cone height is hc = Ht h, where Ht is the total cyclone height. Cyclone C1 and B2 are identical. c
Cyclone C2 and cyclone B4 are equal in Ht /D but they are different in both h/D and hc /D.
A grid independency study has been performed for the tested cyclones. Three different grids have been tested for each cyclone, to be sure that the obtained results are mesh independent. For example, meshes with respectively 490164, 714029 and 1174029 cells have been used for cyclone B2. The computational results on the three grids are presented in Table 7.4. As it can be seen the maximum difference between the results is less than 5%, so the results on the 490164 cells grid can already be considered as sufciently accurate. However, for excluding any uncertainty, computations have been performed using the 714029 cells grid, where the total number of grid points was not that critical with respect to the computation overhead [10, 52]. Figure 7.1(b) shows the surface grid of cyclone C1(B2) used in this study. Table 7.3 gives the total number of cells used for each cyclone after the grid independency study. The hexahedral grids were generated using the GAMBIT grid generator.
Table 7.2: The boundary conditions Boundary Condition Inlet Velocity inlet Outlet Outow Cone tip Wall (no-slip) Other surfaces Wall (no-slip)
129
B1 688170
B2 (C1) 714029
B3 712183
B4 786865
C2 770556
C3 820362
The total number of hexahedral cells after the grid independency study
7.3 Results
The ow velocity can be decomposed into three components. The tangential and the axial velocity components are the major velocity components in comparison with the radial velocity component. Xiang and Lee [186] stated that the tangential velocity is the dominant gas velocity in gas cyclones, which results in the centrifugal force for particle separation. The axial component is responsible for the two ow streams (downward and upward).
7.3.1
Figures 7.2 - 7.4 present the radial proles of the time-averaged tangential and axial velocity and static pressure at nine axial stations (cf. Table 7.5). As expected, the tangential velocity proles exhibit the so-called Rankine vortex, which consists of two parts, an outer free vortex and an inner solid rotation in the center (Fig. 7.2). The tangential velocity distribution in the inner region is rather similar at different sections for the same cyclone. In the outer region, due to the sharp drop in velocity magnitude in the near wall region, the distribution is different but the maximum tangential velocity is similar at all sections for the same cyclone. The axial velocity proles at nine different stations are shown in Fig. 7.3. All cyclones show the inverted W prole but the central dip decreases with increasing the
Table 7.4: The details of the grid independence study for cyclone C1 (B2)
Number of cells 490164 714029 1174029 % differencec
a
2 1 Euler number is the dimensionless pressure drop Eu = P/( 2 Vin ) where P is the static pressure drop, is the gas density, Vin is the gas inlet velocity.
b The Stokes number based on the cut-off diameter; Stk50 = p x2 50 Vin /(18D ) [37]. It is the ratio between the particle relaxation time; p x2 50 /(18) and the gas ow integral time scale; D/Vin where p is the particle density =860 kg/m3 , is the gas viscosity. c
130
7.3. Results
Table 7.5: The position of different plotting sections Section z`/Da
a
S1 2.75
S2 2.5
S3 2.25
S4 2
S5 1.75
S6 1.5
S7 1.25
S8 1.0
S9 0.75
barrel (or cone) height (Except at the cyclone bottom for cyclones C2 and C3, where the axial velocity shows an inverted V prole). The radial proles of the time averaged static pressure are given in Fig. 7.4. Like for the tangential velocity, the axial variations of the static pressure are very small for the same cyclone. The variations become negligible with increasing barrel (or cone) height. Furthermore, the maximum value of the static pressure decreases when the barrel height is increased.
B1
10
10
4
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-2 -0.015
-0.01
-0.005
0.005
0.01
0.015
0 -0.015
C2
12
12
B3
10
10
4
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-2 -0.015
-0.01
-0.005
0.005
0.01
0.015
0 -0.015
C3
12
12
B4
10
10
4
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-2 -0.015
-0.01
-0.005
0.005
0.01
0.015
0 -0.015
Figure 7.2: The radial prole for the time-averaged tangential velocity at different sections. Note: C1=B2.
132
7.3. Results
C1 (B2)
8
S1 S2 S3 S4 S5 S6 S7 S8 S9
8
B1
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
C2
8
S1 S2 S3 S4 S5 S6 S7 S8 S9
8
B3
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
C3
8
S1 S2 S3 S4 S5 S6 S7 S8 S9
8
B4
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
Figure 7.3: The radial prole for the time-averaged axial velocity at different sections. Note: C1=B2.
133
B1
150
200
100
50
150
100
S1 S2 S3 S4 S5 S6 S7 S8 S9
50
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-50 -0.015
-0.01
-0.005
0.005
0.01
0.015
-100 -0.015
C2
200
300 250
B3
150
200
100
50
150
100
S1 S2 S3 S4 S5 S6 S7 S8 S9
50
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-50 -0.015
-0.01
-0.005
0.005
0.01
0.015
-100 -0.015
C3
150
300 250
B4
100
Static pressure (N/m2)
2
200
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
150
100
50
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-50 -0.015
-0.01
-0.005
0.005
0.01
0.015
-100 -0.015
Figure 7.4: The radial prole for the time-averaged static pressure at different sections. Note: C1=B2.
134
7.3. Results
height (either barrel or cone) decreases the pressure drop, the maximum tangential velocity (vortex strength) and the dip in the axial velocity prole. The effect of increasing the cone height on the axial velocity prole is predominant with respect to the barrel height. The swirling motion of the gas generates a strong radial pressure gradient, the pressure being low in the centre of the vortex and high at the periphery. As the strongly swirling gas enters the connes of the vortex nder on its way out of the cyclone, the swirl is attenuated through friction with the wall. This means that further up the vortex nder the pressure in the centre is higher than at the exit of the separation space: a reverse pressure gradient is present [78] as is clear from Fig.7.5. This drives an axial ow with dip in the centre of the vortex nder (inverted W prole); this core ow prevails throughout the entire separation space of the cyclone in spite of the attenuation of swirl in the conical part of the cyclone. With increasing the cone height the pressure distribution becomes atter consequently the pressure force causes the dip in the axial velocity at the center line becomes less and less. That may explain also the change of the axial velocity from cyclones C1 to C3, Figs. 7.5 and 7.6.
250
200
200
150
150
100
100
50
50
0
-50
-50 -0.015
-0.01
-0.005
0.005
0.01
0.015
-100 -0.015
14 12 10
Tangential velocity (m/s)
12
10
8 6 4 2 0 -2 -0.015
-0.01
-0.005
0.005
0.01
0.015
0 -0.015
-2
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
Figure 7.5: Comparison between the radial proles for the time averaged static pressure, tangential and axial velocity at section S6.
136
7.3. Results
ences between the four cyclones are small. The axial velocity patterns for the four cyclones have the shape of an inverted W prole.
Figure 7.6: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From left to right : the static pressure (N/m2 ), the tangential velocity (m/s) and the axial velocity (m/s). From top to bottom cyclone C1-C3.
138
139
7.3. Results
Figure 7.7: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From top to bottom: the static pressure (N/m2 ), the tangential velocity (m/s) and the axial velocity (m/s). From left to right cyclone B1 - B4.
nearly at between cyclones B3 and B4 due to the small changes in both the axial and the tangential velocity prole between the two cyclones. Both Ramachandran [139] and Iozia [84] models agree with the CFD results in the trend of decreasing both the Euler number and Stokes number with increasing barrel height, but differ in slope and values.
6 Euler number (CFD) Euler number (Ramachandran model) Stk50 x 10 3 (CFD) 3 Stk50 x 10 (Iozia model) 2.4 2.2 2 5 1.8 1.6 4.5 1.4 1.2 1 3.5 0.8 0.6 3 3.5 4
5.5
Euler number
H t /D
4.5
Figure 7.8: The effect of barrel height on the Euler number and the Stokes number.
Table 7.6: The Euler numbers and Stokes numbers for cyclones B1-B4 Cyclone h/D hc /h Ht /D Eu Eu (Ramachandran model [139]) Stk50 x 103 Stk50 x 103 (Iozia model [84]) B1 1.0 2.5 3.5 4.39 5.71 1.32 1.94 B2 1.5 1.666 4 3.654 4.77 1.2 1.82 B3 2.0 1.25 4.5 3.33 4.17 1.01 1.72 B4 2.5 1.0 5.0 3.09 3.73 0.95 1.64
Figure 7.9 and Table 7.7 show a sharp decrease of both the Euler number and the Stokes number with increasing the cone height until hc /D=3.3 and a gradual decrease when 3.3 > hc /h > 4.0. This behavior can be explained as follows. As the inlet section is the same in all cyclones, the pressure drop in the inlet section may not vary with increasing the cone height. 140
Stk50 x 10 3
7.3. Results
The sharp decrease in the Euler number between cyclones C1 and C2 is due to the huge decrease in the pressure drop in the cyclone body due to the drop in the maximum tangential velocity (the decay of the vortex strength). There are two competing contributions, increase of the pressure drop due to friction and decrease of the pressure drop due to the reduction in the vortex strength. At the beginning, the decay in the pressure drop due to the decay of the vortex strength overrides the effect of increasing the pressure drop due to wall friction for longer cyclones. This explains the small variation of the Euler number with the cone height for hc > 4.0, which is clear from Fig. 7.5 where the maximum tangential velocities of cyclones C2 and C3 are very close. The trend of decreasing Stokes number with increasing cone height is quite reasonable, as more separation space exists, and the possibility of particles to be captured increases. Although, the vortex strength decreased with a small amount - due to the reduction in the maximum tangential velocity- we estimate that the main contribution to the collection efciency comes from the increased collecting surface with increasing the cone height. The reason of nearly constant Stokes number after Ht /D =5.5 is the change in the axial velocity prole. Figure 7.5 shows a higher kinetic energy of the ow at the cyclone bottom for cyclone C3, that will enhance re-entrainment of some of the captured particles to escape with the upward ow. Consequently low collection efciency and higher cut-off diameter (Stokes number). Because, the differences between the axial and tangential velocity prole between cyclones C2 and C3 are limited.
Table 7.7: The Euler numbers and Stokes numbers for cyclones C1-C3 Cyclone hc /h hc /D Ht /D Eu Stk50 x 103 C1 1.666 2.5 4 3.654 1.2 C2 2.333 3.5 5 2.749 0.465 C3 3.0 4.5 6 2.584 0.315
1.5 3.5 4 4
2 4.5
hc /h H t /D
5
2.5 5.5
3 6
Euler number Stk50 x 10 3
3.5 6.5 4
3.5
3.5 3
Euler number
2.5
2 1.5 1
1.5 0.5 1 0
2.5
3.5
4.5
hc /D
Figure 7.9: The effect of cone height on the pressure drop (Euler number) and cut-off diameter (Stokes number).
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
Figure 7.10: Comparison between the radial proles for the time averaged static pressure, tangential and axial velocity at section S9.
142
Stk50 x 10
2.5
7.3. Results
total height either by increasing the barrel or cone height. The effect of changing the cone height is more important than that of changing the barrel height (cf. Table 7.6, 7.7 and Fig. 7.11). The effect of increasing the ratio of cone to barrel height hc /h on the cyclone performance depends on the dependent variable. If hc /h increases at constant cone height, the Euler number increases linearly with decreasing the barrel height. If hc /h increases at constant barrel height, the Euler number decreases with increasing the cone height. The effect of changing hc /h at constant cone height has a negligible effect on the Stokes number. Increasing hc /h at constant barrel height decreases the Stokes number. This effect becomes negligible for hc /h > 2.75 (Fig. 7.11). This behavior can be explained by inspecting the variation of the time averaged static pressure, tangential and axial velocity proles with cone and barrel height, Fig. 7.12. As is clear from Fig. 7.12 the differences between the maximum tangential velocity in cyclones C2 and C3 are negligible. Furthermore, the effect of the cone height on the ow eld is more signicant than that of the barrel height.
143
Euler number
2.5 2 1.5 1
H t /D
6 Euler number (Barrel height) Euler number (Cone height) 3 Stk50 x 10 (Barrel height) 3 Stk50 x 10 (Cone height) 6
Euler number
0 0.75
1.25
1.5
1.75
2.25
2.5
2.75
0 3.25
hc /h
Figure 7.11: The variation of the Euler number and the Stokes number with the barrel and cone height.
144
Stk50 x 10 3
Stk50 x 10 3
7.3. Results
300 300
250
250
200
200
150
150
100
100
50
50
-50
-50
-100 -0.015
-100 -0.015
12
12
10
10
-2 -0.015
-2 -0.015
-2
-2
-4 -0.015
Figure 7.12: The radial prole for the time averaged static pressure, tangential and axial velocity at three different sections for the six cyclones. From top to bottom : static pressure, tangential and axial velocity. From left to right: S6 - S9. Note: h/D = 1.5 also represents hc /D = 2.5
145
7.4 Conclusions
Six cyclones of different barrel and cone height have been simulated using the Reynolds stress model (RSM), to study the effect of cyclone height on the performance and ow pattern. The following conclusions have been obtained. The maximum tangential velocity in the cyclone decreases with increasing the cyclone (barrel or cone) height. No acceleration occurs in the cyclone space (the maximum tangential velocity nearly constant throughout the same cyclone). Increasing the barrel height, makes a small change in the axial velocity. Increasing the cyclone barrel height decreases the pressure drop and the cut-off diameter. The changes in the performance beyond h/D = 1.8 are small. Increasing the cone height makes a considerable change in the axial velocity. Both the pressure drop and the cut-off diameter decrease with increasing the cyclone cone height. The performance improvement stops after hc /D = 4.0 (Ht /D = 5.5). The effect of changing the barrel height is less signicant on the performance and the ow pattern in comparison with the effect of the cone height.
146
Chapter 8 Optimization
8.1 Introduction
Mathematical optimization refers to the selection of a best element from some set of available alternatives. In the simplest case, this means solving problems in which one seeks to minimize or maximize a real function by systematically choosing the values of real or integer variables from within an allowed set. Generally, the use of the word optimization implies the best result under the circumstances [176]. In 1951, Stairmand [166] presented one of the most popular design guidelines for the high-efciency cyclone separators [147]. Stairmand presented the geometrical ratios for the seven geometrical parameters as: a/D = 0.5, b/D = 0.2, Dx /D = 0.5, Ht /D = 4.0, h/D = 1.5, S/D = 0.5 and Bc /D = 0.375. These values have been obtained based on the Stairmand model [165] which suffers from many shortages [50]. In the Stairmand model [165], the velocity distribution has been obtained from a moment-of-momentum balance, estimating the pressure drop as entrance and exit losses combined with the loss of static pressure in the swirl. The main drawbacks of the Stairmand model are: (1) neglecting the entrance loss by assuming no change of the inlet velocity occurs at the inlet area; (2) assuming a constant friction factor; (3) the effect of the particle mass loading on the pressure drop is not included. All these drawbacks are overcome in the Muschelknautz method of modeling (MM) [77] introduced by Muschelknautz and Trefz [116, 117]. The main benet of MM over other models is its ability to take the following effects into account: a) wall roughness due to both the physical roughness of the materials of construction and to the presence of collected solids. b) the effect of the mass loading and the Reynolds number on cyclone performance. c) the change 147
Chapter 8. Optimization
of ow velocity throughout the cyclone [77]. Consequently, the optimization procedure using a data obtained from MM model via response surface methodology will result in better results than the simple Stairmand model. An alternative approach is using the available experimental data in literature using the articial neural networks (ANNs) approach for the optimization process. The third source of data is the CFD simulations to train the ANNs. A general unconstrained optimization problem can be expressed as [182]: min subject to fi (x), x = [x1 , x2 , . . . , xn ]T , x X where x = [x1 , x2 , . . . , xn ]T is the vector of design variables (cyclone dimensions) with total number of n, X denotes the design space. The fi (x) stands for objective functions with total number of I . I should be larger than or at least equal to 1, which suggests the multi-objective or single objective optimal problem. In Eq. 8.1, f1 (x) is the Euler number and f2 (x) is the cut-off diameter. The design variables are seven geometrical parameters, namely, Dx /D, a/D, b/D, Ht /D, h/D, S/D and Bc /D. The side constraints are: 0.3 Dx /D 0.65, 0.2 a/D 0.65, 0.15 b/D 0.3, 3 Ht /D 6, 0.75 h/D 2, 0.4 S/D 0.75 and 0.1 Bc /D 0.45. There are numerous optimization techniques that can be used for the geometry optimization in the cyclone separator. However, only two techniques have been used in this thesis (Nelder-Mead technique and the genetic algorithms (GA)) for robustness, and availability in the used software. In cases of multi-objective optimization studies, two approaches have been applied, the desirability function and the non-sorted dominated genetic algorithm (NSGA-II) techniques have been selected due to robustness, and availability. More details about the used optimization techniques are given in appendix B. A detailed literature review for the previous optimization studies is given in Sec. 2.5.6, page 21. The present study is an attempt to obtain new optimized cyclone separator designs based on the MM model, experimental data sets and CFD simulations data sets and to investigate the effect of each cyclone geometrical parameter on the cyclone performance using response surface methodology (RSM) and CFD simulations. Table 8.1 summaries the conducted optimization studies. 148 i = 1, 2, . . . , I (8.1)
8.2 Optimization of the Euler number using MM model and Nelder-Mead technique
The present study aims to obtain a new optimized cyclone separator for minimum pressure drop based on the MM model and to investigate the effect of each cyclone geometrical parameter on the cyclone performance using CFD simulations. This section is an extension to the sensitivity analysis study presented in Sec. 4.1, page 45.
Dx
Le
h Li D b
Ht
Bc
Table 8.2 gives the optimum values for cyclone geometrical parameters for minimum pressure drop estimated by MM using the downhill simplex optimization technique available in Statgraphics XV software. 149
Chapter 8. Optimization
Table 8.2: The geometrical parameters for minimum pressure drop using MM model
Factor a/D b/D Dx /D Ht /D h S Bc Low 0.5 0.14 0.2 3 1.0 0.4 0.2 High 0.75 0.4 0.75 7 2.0 2 0.4 Optimum 0.618 0.236 0.622 4.236 1.618 0.620 0.382
8.2.1
Numerical settings The air volume ow rate Qin =0.08 m3 /s for the two cyclones (inlet velocity for Stairmand design is 19 m/s and 13.1 m/s for the new design), air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 P a s. The turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [75]. Velocity inlet boundary condition is applied at inlet, outow at the gas outlet and wall boundary condition at all other boundaries. The nite volume method has been used to discretize the partial differential equations of the model using the SIMPLEC (Semi-Implicit Method for Pressure-Linked Equations-Consistent) method for pressure velocity coupling and QUICK scheme to interpolate the variables on the surface of the control volume. The implicit coupled solution algorithm was selected. The unsteady Reynolds stress turbulence model (RSM) was used in this study with a time step of 0.0001 s. The residence time (cyclone volume/gas volume ow rate) of the two cyclones are close ( 0.25 s). The grid renement study shows that a total number of about 134759 hexahedral cells for Stairmand cyclone and 154746 hexahedral cells for the new design are sufcient to obtain a grid-independent solution, and further mesh renement yields only small, insignicant changes in the numerical solution. These simulations were performed on an eight nodes CPU Opteron 64 Linux cluster using Fluent commercial software. The geometrical values are given in Table 8.3 for the two cyclones (cf. Fig. 8.1).
Table 8.3: The values of geometrical parameters for the two designs (D=0.205 m)
Cyclone Stairmand design New design a/D 0.5 0.618 b/D 0.2 0.236 Dx /D 0.5 0.622 Ht /D 4 4.236 h/D 1.5 1.618 S/D 0.5 0.620 Bc /D 0.36 0.382 Li /D 1.0 1.0 Le /D 0.618 1.618
150
S1 2.75
S2 2.5
S3 2.25
S4 2.0
S5 1.75
S6 1.5
S7 0.25
Sections S1S5 are located in the conical section, section S6 at the cylindrical part and S7 located through the inlet section. b z measured from the inlet section top
Chapter 8. Optimization
Figure 8.2: The contour plots for the time averaged ow variables at sections Y=0 and S7. From top to bottom: static pressure [N/m2 ], tangential velocity [m/s] and axial velocity [m/s]. From left to right Stairmand design and new design respectively.
152
before, including a quasi-forced vortex in the central region and a quasifree vortex in the outer region. The maximum tangential velocity may reach twice the average inlet velocity and occurs in the annular cylindrical part. The tangential velocity distributions for the two cyclones are nearly identical in pattern and values (dimensionless), with the highest velocity occurring at 1/4 of the cyclone radius for both cyclones. This implies a nearly equal collection efciency for both cyclones, as the centrifugal force is the main driving force for particle collection in the cyclone separator. The axial velocity proles for the two cyclones are also very close, exhibiting a M letter shape (also known as inverted W axial velocity prole in some other literatures (cf. Horvath et al. [79])). Part of the ow in the central region moves downward in the two cyclones. This phenomena has been shown in the axial velocity pattern in other published articles [e.g., 79, 159].
The DPM results In order to calculate the cut-off diameters of the two cyclones, 104 particles were injected from the inlet surface with a velocity equals the inlet gas velocity. The particle density p is 860 kg/m3 . The grade efciency curves for the two designs are plotted in Fig. 8.5. The DPM analysis results and the pressure drops for the two cyclones are depicted in Table 8.5. An acceptable agreement between the CFD results and the MM mathematical model has been obtained. While the difference between the two cyclone cut-off diameters is small, the saving in pressure drop is considerable (nearly half the value of Stairmand cyclone). Based on the ow pattern analysis and the DPM results. One can conclude that the cyclone collection efciency for the two cyclones should be very close, with the advantage of low pressure drop in the new design. The authors want to emphasis that only small changes in the geometrical dimensions of the two designs led to this improvement in the performance. 153
Chapter 8. Optimization
8.2.2
Conclusions
Both mathematical modeling (the Muschelknautz method of modeling (MM)) and CFD investigation have been used to understand the effect of the cyclone geometrical parameters on the cyclone performance and a new optimized cyclone geometrical ratios based on MM model has been obtained. The new cyclone design is very close to the Stairmand high efciency design in the geometrical parameter ratio, but superior in low pressure drop at nearly the same cut-off diameter. The new cyclone design results in nearly one-half the pressure drop obtained by the old Stairmand design at the same volume ow rate.
154
Figure 8.3: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S1S3 . From top to bottom: section S1S3. From left to right: time-averaged static pressure, tangential velocity and axial velocity respectively.
155
Chapter 8. Optimization
Figure 8.4: The radial prole for the timeaveraged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S4S6 . From top to bottom: section S4S6. From left to right: time-averaged static pressure, tangential velocity and axial velocity respectively.
156
0.75
0.5
0.25
10
-1
10
10
Figure 8.5: The grade efciency curves for the two designs
157
Chapter 8. Optimization
8.3 Optimization of the Euler number using RBFNN and Nelder-Mead technique
Recently, Articial neural networks (ANNs) have been widely applied in the elds of modeling, prediction, fault detection and process control. In the eld of performance evaluation for cyclone separators, unfortunately, ANNs have not been paid enough attention on their algorithmic advantages [197]. There are few articles about the application of neural network in the eld of cyclone separator. There are six main objectives of this study. (1) Application of the articial neural network to model the pressure drop using experimental dataset. (2) Investigation of the effect of the seven geometrical parameters on the pressure drop based on the trained RBFNN. (3) A detailed comparison between the experimental pressure drop values and the estimated values obtained from different mathematical models. In order to recommend the best mathematical model for future use. (4) Application of the response surface methodology to study the effect of each geometrical parameter on the pressure drop and test the interaction between these parameters using the trained RBFNN. (5) Obtaining the optimum design (geometrical ratios) for minimum pressure drop. (6) CFD study of the new cyclone separator and compare its performance and ow pattern with the Stairmand design.
8.3.1
Why RBFNN Radial basis function neural networks (RBFNNs) are powerful and interesting networks due to their rapid training, generality and simplicity [16]. Girosi and Poggio [65] and Hartman and Keeler [72] proved that RBFNNs are universal approximators and can approximate any continuous function with arbitrary accuracy. Training of these networks is very fast, and they are very good at interpolation [190]. Niros and Tsekouras [120] stated that radial basis function neural networks (RBFNN) have certain advantages over other types of neural networks including better approximation capabilities, simple network structure, and faster learning. Zhao and Su [197] tested three different types of articial neural network to model the pressure drop in cyclone separators, viz. the back propagation neural network (BPNN), the radial basis function neural network (RBFNN) and the generalized regression neural network (GRNN). They stated that compared with the BPNN and GRNN, the RBFNN provides superior prediction performance criteria, better capability of approximation and high robustness. 158
In this study, the radial basis function neural network (RBFNN) is employed to model the pressure drop in the cyclone separator. The structure of RBFNN The radial basis function neural network (RBFNN) is a kind of 3-layered forward network with multi inputs and multi outputs. The rst layer has (m) inputs, while the second is a hidden layer with (L) units, and the third layer has (n) outputs, Fig. 8.6. The transformation function from the inputs to the hidden units, varied radial basis functions (RBF) on different occasions, is nonlinear, whereas the mapping of the hidden units to the outputs is linear [180]. Theoretically, RBFNN has the ability of approaching nonlinear mapping arbitrarily, fr : Rm Rn , as dened by:
L
y = fr (x) = W0 +
i=1
Wi ( x ci )
(8.2)
where x Rm is the input vector, y Rn is the output vector, ci Rm (i = 1, 2, L) is the center vector, is the Euclidian norm, Wi Rm (i = 1, 2, L) is the weight vector, W0 Rm is the bias vector, and () is the Gaussian function, a non negative and nonlinear function with radial symmetry and attenuation versus center, in the form of: (v ) = exp(v 2 /2 2 ) where is a width constant (spread factor ) [180]. In order to comprehensively compare the model performance, the evaluation parameters, normalized mean squared error E 2 and correlation coefcient R are employed as follows [197]: E2 = 1 n 2 (yN i y N i ) n i=1 (8.4) (8.3)
where n is the number of test cases, y is the actual variable, y is the RBFNN output variable and N i is the neurons number of input layer in the RBFNN. N i y N n i=1 (yN i y N ) y (8.5) R= 2 2 n ( y y n y ) y N i N i N N i=1 i=1 159
Chapter 8. Optimization
Input layer
Hidden layer
Output layer
W0 x1
y1
x2
y2
xm
yn Wi
Figure 8.6: Schematic diagram for the radial basis function neural network
Modeling procedures
The ANN modeling procedure can be divided into the following steps: selecting the variables, dividing the sample, optimizing parameters, training and testing simulation and evaluating performance [197]. In this study, all calculations are carried out on a MacBook pro laptop with the hardware congurations: processor, Intel Core 2 Duo (2.4 GHz); memory, 4.0 GB (DDR3-1067 2G 2); hard drive, 320 GB (7200 rpm); with Mac OS X 10.5 system. The Euler number The pressure drop across a cyclone separator essentially depends on the dimensions and operating conditions. Generally, it is proportional to the average dynamic pressure at the inlet and is often dened as [197] P = Eu
2 1 2 g Vin
(8.6)
where Eu is Euler number (the dimensionless pressure drop also called the pressure drop coefcient [197]). The Euler number is a complex nonlinear function of the cyclone geometrical dimensions and is not affected by operating conditions in the high Reynolds number ( Re > 5E4) [50, 77]. The Euler number will be constant for any cyclone conguration regardless of size as long as the dimension ratios remain the same, although the pressure drop varies with different operating conditions (due to the effect 160
of g and Vin ). Therefore, pressure drop can be established by determining experimentally or theoretically for a particular cyclone design and also be modied by the semi-empirical correlations to take the effect of solid loading [197]. In order to determine the Euler number more accurately, all eight dimensions of the cyclone are selected to establish the ANN models because they have the effect on the Euler number to different extent [139, 197]. Usually, these dimensions can be characterized by the barrel diameter D and expressed as seven dimensionless geometric ratios [197]: Eu = f D x a b S Ht h B c , , , , , , D D D D D D D (8.7)
According to Eq. 8.7, seven independent dimensionless geometrical variables and one dependent variable (the Euler number of the cyclone) are selected as respectively the input and output parameters in the ANN model, as presented in Table 8.6. For simplicity, the division of each factor by the barrel diameter D will be dropped. A dataset of 98 samples obtained from the measurements of pressure drop for different cyclone designs available in the literature [42, 139, 197] is used in the present investigation to evaluate the prediction performance of the ANN models. Table 8.7 presents more details about the used dataset including the minimum, mean, maximum and range of the seven dimensionless geometrical ratios. Due to the large difference in the order of magnitude of the value (cf. Table 8.7), the available dataset is transformed into -1 to 1 interval using the Matlab intrinsic function; mapminmax in order to avoid solution divergence [197]. The ANN calculations have been performed using the neural network toolbox available from Matlab commercial software 2010a. Descriptive statistical parameters of the input dataset Table 8.8 shows Pearson product moment correlations between each pair of variables. These correlation coefcients range between -1 and +1 and measure the strength of the linear relationship between the variables. Moreover, shown in parentheses is the P-value which tests the statistical signifTable 8.6: The input and output variables for ANN model Variables Specication Input parameters X1 X2 X3 X4 Dx a b S X5 Ht X6 h X7 Bc Output diameter y Eu
161
Chapter 8. Optimization
Table 8.7: Descriptive statistical parameters for the training dataset
Variable Minimum Mean Maximum Range Dx 0.25 0.429 0.667 0.417 a 0.113 0.630 1.0 0.887 b 0.067 0.211 0.4 0.333 S 0.39 0.891 3.052 2.662 Ht 1.158 3.283 10.97 9.812 h 0.501 1.189 3.5 2.999 Bc 0.14 0.342 1.0 0.86
icance of the estimated correlations. P-values below 0.05 indicate statistically signicant non-zero correlations at the 95.0% condence level. The following pairs of variables have P-values below 0.05: (1) Dx with a and Bc ; (2) a with b, Ht and Bc ; (3) b with Ht and Bc ;(4) S with Ht , h and Bc ; (5) Ht with h and Bc ; (6) h with Bc . These conclusions can be obtained also from Fig. 8.7. From this analysis, however there are some correlations between the input variables, but this dataset is still reliable. The same dataset has been used successfully by other researchers [e.g., 194, 197]. Furthermore, it is the only available experimental data set in the literature. K-fold cross validation For the calibration of the RBFNN, the spread factor plays an important role in the regression model. To obtain the optimum value for a multistep search technique is used. In the multi-step search technique, the seeking for the optimum value is performed in two steps [197]. The rst step is a coarse search to determine the best range of values. The second search is a ne search in the best range. To avoid overtting, the K-fold cross validation has been employed [8, 80, 197]. The original sample of data is randomly portioned into K subsamples. A single subsample is used for the validation (testing) and the other remaining K-1 subsamples are used for the training. The process of training and testing is then repeated for each of the K possible choices of the subset omitted from the training. The average performance on the K omitted subsets is then our estimate of the generalization performance. This procedure has the advantage that is allows us to use a high proportion of the available data (a fraction 1 1/K) for training, while making use of all the data points in estimating the
Table 8.8: Correlations between each pair of input variables
Dx a Dx -0.377 (0.000) a -0.377 (0.000) b - 0.121 (0.234) 0.442 (0.000) S 0.092 (0.367) 0.148 (0.145) Ht 0.189 (0.061) 0.288 (0.004) h 0.096 (0.346) -0.052 (0.610) Bc 0.199 (0.049) 0.270 (0.007) The bold value shown in parentheses is the P-value. b -0.121 (0.234) 0.442 (0.000) 0.170 (0.093) 0.239 (0.017) 0.194 (0.054) 0.243 (0.015) S 0.092 ( 0.367) 0.148 (0.145) 0.170 (0.093) 0.378 (0.000) 0.685 (0.000) 0.526 (0.000) Ht 0.189 (0.061) 0.288 (0.004) 0.239 (0.017) 0.378 (0.000) 0.393 (0.000) 0.555 (0.000) h 0.096 (0.346) -0.052 (0.610) 0.194 (0.054) 0.685 (0.000) 0.393 (0.000) 0.470 (0.000) Bc 0.199 (0.049) 0.270 (0.007) 0.243 (0.015) 0.526 (0.000) 0.555 (0.000) 0.470 (0.000)
162
Figure 8.7: Qualitative representation of the correlations between each pair of input variables. For uncorrelated pair of variables, the data will be well distributed and no linear correlation can be obtained e.g., Dx and b (rst row, third column). The high correlation between Ht and h is clear (fth row, sixth column), where one can t easily t a straight line.
generalization error. The disadvantage is that we need to train the network K times. Typically K = 10 is considered reasonable and most widely used [18]. For the radial basis function neural networks (newrb in Matlab 2010a), the learning process is a must to obtain the weights (the width of the radial basis function units). In order to obtain the optimum value of the spread factor , the multi-step search technique with 10-fold cross validation from the interval of (0 - 1) with the performance goal of 1E-5, the maximum number of neurons in the hidden layer equals the training sample size of 98 neurons, the number of neurons to add between displays is 10. Based on the mentioned settings, the optimum value of equals 0.191 (E 2 =1.1321E-06), this value is different than that obtained by Zhao and Su [197] ( equals 0.32 (E 2 =5.84E-04)) the reason can be referred to the lower goal used in the current study (the goal used in Zhao and Su [197] was 1E-4, and all other settings are identical). 8.3.1.1 Fitting the ANN Table 8.9 presents more details about the validation of the used RBFNNs. Both the average, minimum, maximum and range of the input (Euler number) and the predicted Euler number are given. It is clear from Table 8.9 that the ANNs preserved the descriptive statistical parameters of the in163
Chapter 8. Optimization
Table 8.9: Validation of the used RBFNN
Experimental x y 23.268 23.268 2.3 1.745 155.3 155.985 153.0 154.24 0.999 1.311E-4 0.017 0.999 MM x 15.150 1.34 138.0 136.66 y 15.150 1.164 137.235 136.071 0.996 1.212E-4 0.012 0.999 Stairmand x y 20.206 20.206 2.88 2.892 132.0 133.521 129.12 130.629 0.999 9.185E-5 0.011 0.999 Ramachandran x y 22.543 22.543 1.85 1.793 153.0 150.854 151.15 149.061 0.999 1.442E-4 0.020 0.999 Shepherd x y 17.774 17.774 0.957 1.197 92.2 92.543 91.243 91.346 0.999 5.411E-5 0.005 1.0
Average Minimum Maximum Range Correlation Coefcient, R Mean squared error, E 2 Intercept Slope
x is the input to the RBFNN and y is the predicted value. Both x and y represent the Euler number. The values of R, E 2 , intercept and slope are that for the testing stage.
put data. The correlation coefcient between the input and the output and the mean squared error are given for each RBFNN. The intercept and the slope of the adjusted line between the input and the predicted value of the ANN are also given. The congured RBFNN predictions versus experimental data and four other models for cyclone Euler number are shown in Fig. 8.8. According to Fig. 8.8, it can be seen that the ANN models are able to attain the high training accuracy. The training mean square errors for the experimental values and the four mathematical models (MM, Stairmand, Ramachandran and Shepherd and Lapple), have the values 1.311E-4, 1.212E4, 9.185E-5, 1.442E-4 and 5.411E-5 respectively (Fig. 8.8). This indicates that, compared with traditional models of curve tting, the models based on articial intelligence algorithm have a superior capability of nonlinear tting. Especially, the RBFNN has its unique and optimal approximation characteristics in learning process [197]. Figure 8.8 illustrates the agreement between the ANN input and output. The obtained relation is a typical linear relation with a coefcient of correlation close to 1 (R > 0.999). The agreement between the input and output of the ANN is also clear from the value of the mean squared error E 2 (< 1.5E 4). That means, the trained neural network predicts very well the Euler number values and can be used in cyclone design and performance estimation. Table 8.9 and Fig. 8.8 present different performance indicators as a validation of the proposed model for experimental values.
8.3.2
In order to evaluate the performance of the four tested mathematical models in comparison with the experimental values, the percentage residual 164
Predicted value
100
50
50
100
150
Input value
150
150
Predicted value
50
Predicted value
50 100 150
100
100
50
50
100
150
Input value
Input value
(b) MM model
(c) Stairmand
100 150 Data point Linear fit 80 Data point Linear fit
Predicted value
Predicted value
50 100 150
100
60
40
50 20
20
40
60
80
100
Input value
Input value
165
Chapter 8. Optimization
400
Residual error %
20
40
60
80
100
Run number
Figure 8.9: Percentage residual error for the four tested models based on 98 dataset.
error (Eq. 8.8) for each model has been plotted in Fig. 8.9. % error = Model value - Experimental value 100 Experimental value (8.8)
Figure 8.9 depicts that MM model underestimate the pressure drop by around 50%. Also the percentage errors for other models are between 50%. The peaks in error are almost the same for all models. The residual error of MM model is almost the lowest. The reasons of these peaks may be due to the high values of geometrical swirl number for some of the available dataset (cf. Ramachandran et al. [139]), where Sg = Dx D/(4a b). Sg varies between 1 and 5 for industrial cyclones [74]. The Sg values in the 98 dataset used in this study varies from 2.18 to 92.67. As the swirl number is a measure for the ratio of tangential to axial momentum [74, 75, 154], the high values of Sg may cause violation of the simplied assumptions used in the models. Figure 8.9 seems to indicate that the MM model among the other mathematical models is best suited for estimation of the pressure drop. The effect of geometrical parameters on the Euler number The effects of the geometrical parameters on the Euler number are depicted in Fig. 8.10. To study the effect of each parameter, the tested RBFNN model has been used by varying one parameter at a time from its minimum to maximum values of the available 98 dataset, while the other parameters are kept constant at their mean values (cf. Table 8.7). Figure 8.10(a) indicates the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b and the total height Ht . Less effect is due to the cylinder height h (for h > 2.5) and the inlet height a (for a > 0.55). 166
The effect of the geometrical parameters on the Euler number obtained from the MM model (Fig. 8.10(b)) is very close to that obtained from the analysis of the experimental dataset but with underestimation of the Euler number. This supports the use of the MM model in the Euler number estimation for cyclone separators [29, 50, 77]. The situation for the Stairmand model is not the same as that for the experimental data or MM model. Here, the effect of many geometrical parameters attenuated. The effect of the inlet width becomes insignicant. That is not realistic, as for high values of inlet width, a considerable part of the incoming ow will impact directly the vortex nder which increases the entrance loss and consequently, the total pressure drop in the cyclone separator. The Ramachandran model was initially constructed by curve tting based on the used 98 dataset (cf., Ramachandran et al. [139] for more details). So it depicts a better agreement with the experimental values of the Euler number than that of Stairmand model. The Ramachandran model predicts almost the same Euler number variation with the inlet width. In the Shepherd and Lapple model only the inlet dimensions (aandb) and vortex nder diameter 2 Dx affect the pressure drop (Eu = 16ab/Dx ) as is clear from Fig. 8.10(e). Figure 8.11 compares the effect of each individual geometrical parameter using predictions with the ANN based on respectively the experimental data and the four tested models. The following conclusions can be drawn from the analysis of Fig. 8.11: Dx : All the models (except Shepherd and Lapple) show the same variation of the Euler number with increasing Dx . For Dx > 0.5 there is an exact matching between the results of MM model and the experimental values. a : Both the Ramachandran and the Shepherd and Lapple models predict a (nearly) linear relation between the inlet height and the Euler number. The effect of changing a predicted by the Stairmand model is in accordance with the experimental values for a > 0.8. b : For small values of b (b < 0.15), the MM model results are in good agreement with the experimental values. Also the trends of all models in this range are matching the trend of the experimental values. Beyond this range, Both the Ramachandran and the Stairmand models agree well with the experimental values trend. S : The trend of the MM model results is similar to that of the experimental values, although there is a shift in the values of the Euler number. The Shepherd and Lapple model does not present any effect of changing S on the Euler number (for Shepherd and Lapple model, 2 Eu = 16ab/Dx ), while the MM model always underestimates the ef167
Chapter 8. Optimization
Ht
6
10
h, S
0.5 1
Dx a b S Ht h Bc
1.5
2.5
3.5
40
Euler number
30
20
10
D x, a, b, Bc
Ht
6
10
Ht
6
10
h, S
0.5 1
Dx a b S Ht h Bc
h, S
2.5 3 3.5 0.5 1
Dx a b S Ht h Bc
1.5
1.5
2.5
3.5
40
40
Euler number
Euler number
0.4 0.6 0.8 1
30
30
20
20
10
10
0 0.2
D x, a, b, Bc
D x, a, b, Bc
(b) MM model
Figure 8.10: The effect of geometrical parameters on the Euler number using the trained neural networks based on experimental dataset and four different mathematical models.
168
Ht
6
10
Ht
6
10
h, S
0.5 1 1.5 2 2.5 3 3.5 0.5 1 1.5 2
h, S
2.5 3 3.5
40
Euler number
20
Euler number
30
Dx a b S Ht h Bc
40
30
Dx a b S Ht h Bc
20
10
10
D x, a, b, Bc
D x, a, b, Bc
Figure 8.10: (continued) The effect of geometrical parameters on the Euler number using the trained neural networks based on experimental dataset and four different mathematical models.
fect of S . Both the Stairmand and the Ramachandran models may over/underestimate its effect on Euler number. Ht : Both the Stairmand and the Ramachandran models do not show a signicant effect of Ht on the Euler number. The trends of both MM and Ramachandran model are almost the same as that for the experimental values. h : For higher values of h all models (except the Shepherd and Lapple model) give nearly the same value of Euler number. Bc : Nearly, all the models (except the Shepherd and Lapple model) show the same trend in the changing the Euler number with the cone tip diameter.
Chapter 8. Optimization
Table 8.10: The values of the independent variables used in the design of experiment Variables Vortex nder diameter, Dx =X1 Inlet height, a=X2 Inlet width, b=X3 Vortex nder length, S =X4 Total cyclone height, Ht =X5 Cylinder height, h=X6 Cone tip diameter, Bc =X7 minimum 0.2 0.4 0.14 0.4 3.0 1.0 0.2 center 0.475 0.55 0.27 1.2 5.0 1.5 0.3 maximum 0.75 0.7 0.4 2.0 7.0 2.0 0.4
individual factors and interactions. Fitting the model Analysis of variance (ANOVA) showed that the resultant quadratic polynomial models adequately represented the experimental data with the coefcient of multiple determination R2 being 0.965843. This indicates that the quadratic polynomial model obtained was adequate to describe the inuence of the independent variables studied [189]. Analysis of variance (ANOVA) was used to evaluate the signicance of the coefcients of the quadratic polynomial models (see Table 8.12). For any of the terms in the models, a large F-value (small P-value) would indicate a more signicant effect on the respective response variables. Based on the ANOVA results presented in Table 8.12, the variable with the largest effect on the pressure drop (Euler number) was the linear term of vortex nder diameter, followed by the linear term of inlet width and vortex nder length (P < 0.05); the other four linear terms (inlet height, barrel height, cyclone total height and cone tip diameter) did not show a signicant effect (P > 0.05). The quadratic term of vortex nder diameter, vortex nder length and cyclone total height also had a signicant effect (P < 0.05) on the pressure drop; however, the effect of the other four quadratic terms was insignicant (P > 0.05). Furthermore, the interaction between the inlet dimensions and vortex nder diameters (P < 0.05) also had a signicant effect on the pressure drop, while the effect of the remaining terms was insignicant (P > 0.05). Analysis of response surfaces For visualization of the calculated factor, main effects plot, Pareto chart and response surface plots were drawn. The slope of the main effect curve 170
Table 8.11: The geometrical dimensions and Euler number for the used cyclones
Exp. No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 New design Stairmand design X1 a/D 0.7 0.55 0.55 0.55 0.7 0.55 0.55 0.55 0.55 0.4 0.55 0.55 0.55 0.7 0.7 0.7 0.4 0.7 0.55 0.55 0.55 0.55 0.55 0.55 0.4 0.4 0.4 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.4 0.55 0.55 0.7 0.55 0.7 0.55 0.55 0.4 0.4 0.7 0.4 0.4 0.55 0.7 0.4 0.55 0.7 0.55 0.4 0.55 0.55 0.7 0.55 0.55 0.55 X2 b/D 0.27 0.27 0.27 0.27 0.4 0.14 0.4 0.27 0.27 0.4 0.27 0.4 0.14 0.4 0.27 0.27 0.4 0.27 0.4 0.27 0.27 0.14 0.4 0.27 0.14 0.27 0.27 0.27 0.27 0.27 0.4 0.4 0.14 0.27 0.14 0.27 0.14 0.27 0.27 0.27 0.27 0.14 0.14 0.27 0.4 0.27 0.27 0.27 0.27 0.27 0.27 0.27 0.14 0.27 0.14 0.27 0.27 0.14 0.27 0.27 0.27 0.27 0.4 0.27 X3 Dx /D 0.75 0.475 0.475 0.2 0.475 0.475 0.475 0.475 0.475 0.475 0.475 0.75 0.2 0.475 0.2 0.475 0.475 0.75 0.475 0.475 0.2 0.75 0.2 0.475 0.475 0.2 0.2 0.475 0.475 0.75 0.475 0.2 0.2 0.75 0.475 0.475 0.475 0.75 0.75 0.475 0.475 0.475 0.75 0.2 0.75 0.2 0.75 0.475 0.475 0.475 0.475 0.475 0.475 0.475 0.475 0.475 0.2 0.475 0.75 0.475 0.475 0.475 0.475 0.475 X4 Ht /D 5 7 3 7 7 5 5 5 3 7 5 5 5 3 5 5 3 5 5 7 3 5 5 5 7 5 5 5 5 7 5 5 5 3 5 5 5 3 5 5 7 3 5 5 5 7 5 5 5 5 5 3 7 5 5 5 3 3 7 3 5 5 5 7 X5 h/D 1 2 2 1.5 1.5 1 1 1.5 1 1.5 1.5 1.5 1.5 1.5 1 1.5 1.5 2 2 1 1.5 1.5 1.5 1.5 1.5 1 2 1.5 1.5 1.5 1 1.5 1.5 1.5 2 1.5 2 1.5 1 1.5 1 1.5 1.5 2 1.5 1.5 2 1.5 1.5 1.5 1.5 2 1.5 1.5 1 1.5 1.5 1.5 1.5 1 1.5 1.5 2 2 X6 S/D 1.2 2 0.4 1.2 1.2 1.2 1.2 1.2 2 1.2 1.2 0.4 0.4 1.2 1.2 0.4 1.2 1.2 1.2 0.4 1.2 0.4 0.4 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 2 2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 2 1.2 2 1.2 2 1.2 1.2 2 2 0.4 0.4 2 1.2 2 1.2 0.4 1.2 1.2 1.2 0.4 2 1.2 1.2 0.4 X7 Bc /D 0.3 0.3 0.3 0.2 0.3 0.4 0.4 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.4 0.3 0.3 0.2 0.3 0.2 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.2 0.2 0.3 0.3 0.4 0.4 0.3 0.2 0.2 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.4 0.3 0.4 0.4 0.2 0.4 0.3 0.3 0.2 0.2 0.2 0.4 0.3 0.4 0.3 0.2 0.3 0.4 0.3 Y1 Euler Number 4.16 6.93 9.30 65.70 11.90 4.34 9.19 7.94 8.87 5.88 7.94 3.24 24.50 13.40 105.00 10.30 7.05 4.07 9.19 7.50 77.50 2.42 130.00 7.94 3.17 41.90 40.70 7.94 7.94 3.35 9.45 128.00 23.80 3.82 4.22 7.94 4.33 3.86 2.99 7.94 7.09 6.03 2.23 103.00 3.05 64.10 2.92 5.70 9.64 6.34 6.12 8.51 4.70 5.88 4.51 10.60 76.30 4.06 3.25 9.77 9.91 7.94 9.01 7.30
0.618 0.5
0.236 0.2
0.618 0.5
4.236 4
1.618 1.5
0.618 0.5
0.3819 0.36
5.24 5.79
The new design based on MM model and downhill simplex optimization scheme (Euler number =5.24, cut-off diameter =1.77m ). The standard Stairmand high efciency cyclone design, (Euler number =5.79, cut-off diameter =1.54m based on MM model).
171
Chapter 8. Optimization
Table 8.12: Analysis of variance of the regression coefcients of the tted quadratic equationa
Variable 0 Linear 1 2 3 4 5 6 7 Quadratic 11 22 33 44 55 66 77 Interaction 12 13 14 15 16 17 23 24 25 26 27 34 35 36 37 45 46 47 56 57 67 R2
a
Regression coefcient 138.604 -485.694 20.2354 325.517 -10.7497 -8.7097 -19.1361 -167.261 515.706 113.392 -13.0018 10.9817 1.04634 6.06181 232.346 -158.258 -345.242 -4.21438 8.42261 31.8848 95.7341 -143.344 20.5571 -3.23107 -16.3805 -6.62067 -26.3282 0.0205865 32.7526 -28.156 -0.62389 -2.06662 -18.6467 -2.60425 2.35394 -5.32565 0.965843
F-Ratio
P-Value
96.89 0.06 59.28 6.87 0.06 0.57 0.60 489.19 2.09 0.02 15.89 5.63 0.74 1.74 7.61 27.22 0.15 3.83 3.43 1.24 1.40 1.09 0.17 0.27 0.00 1.34 0.00 0.81 0.02 0.18 0.12 0.40 1.21 0.04 0.01
0.0000 0.8152 0.0000 0.0140 0.8161 0.4575 0.4464 0.0000 0.1590 0.9017 0.0004 0.0247 0.3974 0.1983 0.0101 0.0000 0.6981 0.0602 0.0744 0.2752 0.2473 0.3060 0.6851 0.6076 0.9668 0.2569 0.9982 0.3758 0.8782 0.6763 0.7294 0.5334 0.2803 0.8437 0.9112
Bold numbers indicate signicant factors as identied by the analysis of variance (ANOVA) at the 95% condence level.
172
is proportional to the size of the effect and the direction of the curve species a positive or negative inuence of the effect [50, 61] (Fig.8.12(a)). Based on the main effect plot, the most signicant factors on the Euler number are: (1) the vortex nder diameter Dx , with a secondorder curve with a wide range of inverse relation and a narrow range of direct relation, (2) the inlet width b linearly related to the Euler number, (3) the cyclone total height Ht , (4) the vortex nder length S , whereas the other factors have an insignicant effect. Pareto charts were used to graphically summarize and display the relative importance of each parameter with respect to the Euler number [50]. The Pareto chart shows all the linear and second-order effects of the parameters within the model and estimates the signicance of each with respect to maximizing the Euler number response. A Pareto chart displays a frequency histogram with the length of each bar proportional to each estimated standardized effect [30]. The vertical line on the Pareto charts judges whether each effect is statistically signicant within the generated response surface model; bars that extend beyond this line represent effects that are statistically signicant at a 95% condence level. Based on the Pareto chart (Fig. 8.12(b)) and ANOVA table (Table 8.12) there are ve signicant parameters (eight terms in the ANOVA table; Table 8.12) at a 95% condence level: the vortex nder diameter Dx , the inlet width b, the total cyclone height Ht , the vortex nder length S and the inlet height a (due to interaction with Dx ). Therefore, the Pareto chart is a perfect supplement to the main effects plot. To visualize the effect of the independent variables on the dependent ones, surface response of the quadratic polynomial models were generated by varying two of the independent variables within the experimental range while holding the other factors at their central values [189] as shown in Fig. 8.13. Thus, Fig. 8.13(b) was generated by varying the inlet height a and the inlet width b while keeping the other ve factors constant. The trend of the curve is linear, with a more signicant effect for the inlet width b, and a weak interaction between the inlet height a and width b. The response surface plots of Figs. 8.13(a), 8.13(c) and 8.13(d) show that there are strong interactions between the vortex nder diameter Dx and respectively the inlet height a, the cyclone total height Ht and the vortex nder length S . The effect of vortex nder length S is less signicant with respect to the vortex nder diameter Dx , but its effect is higher than that of the barrel height h and the cone tip diameter Bc (cf., Fig. 8.12). 173
Chapter 8. Optimization
8.3.3.1 Optimization (Nelder-Mead method) In this optimization problem, the objective function is the Euler number (f1 (x) in Eq. 8.1). Table 8.13 gives the optimum values for cyclone geometrical parameters for minimum pressure drop. It is clear from Table 8.13 that the new optimized design is very close to the Stairmand design in many geometrical parameters, whereas the new ratios will lead to the minimum pressure drop. To understand the effect of this small change in the geometrical ratios on the ow eld pattern and performance, a CFD study for the two designs is needed.
Table 8.13: The geometrical parameters for minimum pressure drop using RBFNN Factor Dx a b S Ht h Bc Low 0.2 0.5 0.14 0.4 3.0 1.0 0.2 High 0.75 0.75 0.4 2.0 7.0 2.0 0.4 Stairmand design 0.5 0.5 0.2 0.5 4.0 1.5 0.375 Optimum 0.487 0.629 0.203 0.733 4.852 1.633 0.383
174
50
Exp. MM Stairmand Ramchandran Shepherd
25
Exp. MM Stairmand Ramchandran Shepherd
40
20
Euler number
30
Euler number
0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
15
20
10
10
0 0.2
0.2
0.4
0.6
0.8
Dx
(a) Dx
(b) a
30
Exp. MM Stairmand Ramchandran Shepherd
65 60 55 50
Exp. MM Stairmand Ramchandran Shepherd
25
20
45
Euler number
Euler number
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
40 35 30 25 20
15
10
15 10
(c) b
(d) S
Figure 8.11: Comparison between the effect of each geometrical parameters on the Euler number using different models and experimental values based on RBFNN.
175
Chapter 8. Optimization
20
Exp. MM Stairmand Ramchandran Shepherd Exp. MM Stairmand Ramchandran Shepherd
30
15
Euler number
20
Euler number
2 4 6 8 10 12
10
10 5
0.5
1.5
2.5
3.5
Ht
(e) Ht
(f) h
40
30
Euler number
20
10
0.2
0.4
0.6
0.8
Bc
(g) Bc
Figure 8.11: (continued) Comparison between the effect of each geometrical parameters on the Euler number using different models and experimental values based on RBFNN.
176
177
Chapter 8. Optimization
(a) Dx versus a
(b) a versus b
(c) Dx versus Ht
(d) Dx versus S
178
The outlet section is above the cyclone surface by Le = 0.618D. The inlet section located at a distance Li = D from the cyclone center.
179
Chapter 8. Optimization
Figure 8.15: The contour plots for the time averaged ow variables at sections Y=0 (cf. Fig.8.1). From top to bottom: Stairmand design and the new design respectively. From left to right: the static pressure (N/m2 ), the tangential and axial velocity (m/s). Note: both cyclones have the same barrel diameter and air volume ow rate.
181
Chapter 8. Optimization
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
2000
0.8
0.6
1500
1.5
0.4
1000
0.2
500
0.5
0 -0.1
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
0.8
2000
0.6
1500
1.5
0.4
1000
0.2
500
0.5
-0.2 0 -0.1 -0.05 0 0.05 0.1 0 -0.1 -0.4 -0.1 -0.05 0 0.05 0.1
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
0.8
2000
0.6
1500
1.5
0.4
1000
0.2
500
0.5
-0.2 0 -0.1 -0.05 0 0.05 0.1 0 -0.1 -0.4 -0.1 -0.05 0 0.05 0.1
Figure 8.16: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S1S3 . From top to bottom: section S1S3. From left to right: time-averaged static pressure, tangential and axial velocity respectively.
182
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
0.8
2000
0.6
1500
1.5
0.4
1000
0.2
500
0.5
-0.2 0 -0.1 -0.05 0 0.05 0.1 0 -0.1 -0.4 -0.1 -0.05 0 0.05 0.1
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
0.8
2000
0.6
1500
1.5
0.4
1000
0.2
0.5
500
-0.2 0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.4 -0.1 -0.05 0 0.05 0.1
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
2000
0.8
0.6
1500
1.5
0.4
1000
0.2
0.5
500
Figure 8.17: The radial prole for the timeaveraged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S4S6 . From top to bottom: section S4S6. From left to right: time-averaged static pressure, tangential and axial velocity respectively.
183
Chapter 8. Optimization
Table 8.15: The position of different sectionsa
Section z`/D
a b b
S1 2.75
S2 2.5
S3 2.25
S4 2.0
S5 1.75
S6 1.5
Sections S1S5 are located in the conical section, section S6 at the cylindrical part. z` is measured from the top of the inlet section (cf. Fig. 8.1).
The velocity eld Based on the contour plots of the timeaveraged tangential velocity, Fig. 8.15, and the radial proles at sections S1S6 shown in Figs. 8.16 and 8.17, the following conclusions can be drawn. The tangential velocity prole at any section is composed of two regions, an inner and an outer one. In the inner region, the ow rotates approximately like a solid body (forced vortex), where the tangential velocity increases with radius. After reaching its peak the velocity decreases with radius in the outer part of the prole (free vortex). This prole is a so-called Rankine type vortex as mentioned before, including a quasi-forced vortex in the central region and a quasifree vortex in the outer region. The maximum tangential velocity may reach twice the average inlet velocity and occurs in the annular cylindrical part. The tangential velocity distribution for the two cyclones are approximately nearly identical in pattern and values (dimensionless), with the highest velocity occurring at 1/4 of the cyclone radius for both cyclones. This implies a nearly equal collection efciency for both cyclones, as the centrifugal force is the main driving force for particle collection in the cyclone separator. The axial velocity proles for the two cyclones are also very close, exhibiting the inverted W axial velocity prole [79]. Part of the ow in the central region moves downward in the two cyclones. This phenomena has been shown in the axial velocity pattern in other published articles [e.g., 79, 159]. The DPM results In order to calculate the cut-off diameters of the two cyclones, 104 particles were injected from the inlet surface with a velocity equals the air inlet velocity. The particle density p is 860 kg/m3 and the maximum number of time steps for each injection was 200000 steps. The DPM analysis results and the pressure drops for the two cyclones are depicted in Table 8.16. Although, the difference between the two cyclone cut-off diameters is small, the saving in the pressure drop is considerable (nearly 25% the value of Stairmand cyclone). Based on the ow pattern analysis and the DPM results, one can conclude that the cyclone collection efciency for the two cyclones is very close, with 184
the advantage of low pressure drop in the new design. The authors want to emphasis that only small changes in the geometrical dimensions of the two designs lead to this improvement in the performance.
8.3.5 Conclusions
In order to predict accurately the complexly non linear relationships between pressure drop and geometrical dimensions, a radial basis neural network (RBFNN) is developed and employed to model the pressure drop for cyclone separators. The neural network has been trained and tested by the experimental data available in literatures. The following conclusions can be drawn from analysis of the obtained results: The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone pressure drop. Four mathematical models (Muschelknautz method MM, Stairmand, Ramachandran and Shepherd and Lapple) have been tested against the experimental values. The residual error of MM model is the lowest. Also, one can multiply the calculated value using MM by 1.5 to get the experimental value (as a rough approximation) The analysis indicates the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b and the total height Ht . The response surface methodology has been used to t a second order polynomial to the RBFNN. The second-order polynomial has been used to get a new optimized cyclone for minimum pressure drop using Nelder-Mead technique. A comparison between the new design and the standard Stairmand design has been performed using CFD simulation with the Reynolds stress model to get a clear vision of the ow eld pattern in the new design. CFD results show that, the new cyclone design is very close to the Stairmand high efciency design in the geometrical parameter ratio, and superior for low pressure drop at nearly the same cut-off diameter. The new cyclone design results in nearly 75% of the pressure drop obtained by the old Stairmand design at the same volume ow rate. 185
Chapter 8. Optimization
Study objectives There are four objectives of this study. (1) Investigation of the effect of the seven geometrical parameters on the cyclone separator performance (the pressure drop and cut-off diameter) based on the experimental data for the pressure drop and the most robust mathematical models for the cut-off diameter. (2) Study the possible interaction between the seven geometrical parameters affecting the cyclone performance using response surface methodology. (3) Multi-objective optimization to obtain new geometrical ratios for optimum performance (minimum pressure drop and minimum cut-off diameter). (4) Obtaining the optimum design (geometrical ratios) of the cyclone separator for minimum pressure drop using the genetic algorithm optimization technique, followed by a comparison of the numerical simulations of the optimal design and the Stairmand design using the Reynolds stress turbulence model. 186
(8.9)
where is the gas viscosity, Q is the gas volume ow rate, p is the particle density. HCS is the core height (height of the control surface, Fig. 8.18), Vmax is the maximum tangential velocity, that occur at the edge of the control surface CS . In this model, the value of core diameter dc , and the tangential velocity at the core edge; Vmax are calculated from regression of experimental data using Eq. A.34. Vmax = 6.1Vin a b/D2
0.61
(Dx /D)
0.25
0.74
(Ht /D)
1.53
0.33
(8.10) (8.11)
(Dx /D)
Chapter 8. Optimization
Figure 8.18: The control surface (core edge) used in the Iozia and Leith model [85].
where Vin is the area-average inlet velocity. The following expression obtained from trigonometry relations and can be used to calculate HCS . (R
Dx 2 )(Ht Bc
2
HCS
= =
R( (Ht S )
h)
+ (h S )
if Bc > Dx if Bc Dx (8.12)
It is clear from this model that the most important geometrical parameters that affect the cyclone cut-off diameter are the vortex nder diameter,the ratio of inlet area to exit area, cyclone height. Based on this model, the cut-off diameter is a function of the inlet gas velocity (i.e., a function of both gas volume ow rate, Barrel diameter, inlet section height and width), gas viscosity and particle density. For this particular study presented in Fig.8.20. The following values have been used: Barrel diameter D =0.1 m, air ow rate =0.8333 l/s, air viscosity 1.0E-5 Pa s and particle density 860 kg/m3 . This means that the obtained results will be valid for this particular case (used for just demonstration). However, the authors believe the variation of the cut-off diameter due to variations of cyclone geometrical dimensions is superior to the effect of these operating parameters, which is quite difcult to cover their range of operating conditions. The cut-off diameter x50 for cyclone separator is always given in units of m. Another way to represent x50 is using a dimensionless number; Stokes number. The Stokes number based on the cut-off diameter; Stk50 = p x2 50 Vin /(18D ) [37]. It is the ratio between the particle relaxation time; p x2 50 /(18) and the gas ow integral time scale; D/Vin . 188
Fitting the ANNs Tables 8.17 and 8.18 present more details about the validation of the used RBFNNs. Both the average, standard deviation, minimum, maximum and range of the input and the predicted values are given. It is clear from Tables 8.17 and 8.18 that the ANNs preserved the descriptive statistical parameters of the input data. The correlation coefcient between the input and the output and the mean squared error are given for each RBFNN. The intercept and the slope of the adjusted line between the input and the predicted value of the ANN are also given. The congured RBFNN predictions versus experimental data for the Euler number and the Iozia and Leith mathematical model for the Euler number and cut-off diameter are shown in Fig. 8.19. According to Fig. 8.19, it can be seen that the ANN models are able to attain the high training accuracy. The training mean square errors for the experimental values and the Iozia and Leith model have the values 1.311E-4 and 3.258E-4 respectively (Fig. 8.19). This indicates that, compared with traditional models of curve tting, the models based on articial intelligence algorithm have a superior capability of nonlinear tting. Especially, the RBFNN has its unique and optimal approximation characteristics in learning process [197]. Figure 8.19 illustrates the agreement between the ANNs input and output. The obtained relation is a typical linear relation with a coefcient of correlation close to 1 (R > 0.999). The agreement between the input and output of the ANN is also clear from the value of the mean squared error E 2 . That means, the trained neural networks predict very well both the Euler number and cut-off diameter values and can be used in cyclone design and performance estimation. Tables 8.17, 8.18 and Fig. 8.19 present different performance indicators as a validation of the proposed model for experimental values. 189
Chapter 8. Optimization
Table 8.18: Validation of the used RBFNN to model the cut-off diameter
Average Minimum Maximum Range Correlation Coefcient, R Mean squared error, E 2 Intercept Slope x 8.219 3.64 15.3 11.66 0.99915 3.258E-4 0.014 0.999 y 8.219 3.716 15.405 11.689
x is the input to the RBFNN and y is the predicted value. Both x and y represent the cut-off diameter. The values of R, E 2 , intercept and slope are that for the testing stage.
14
120
12
10
40 6
4 40 80 120 4 6 8 10 12 14
Figure 8.19: Linear regression of the RBFNNs for the Euler number and the cut-off diameter.
8.4.1.1 The effect of geometrical parameters on the cut-off diameter based on RBFANN The effect of the geometrical parameters on the Euler number based on the trained RBFNN has been presented in Sec. 8.3.2, page 166. The study acknowledge the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b and the total height Ht . Less effect is due to the cylinder height h (for h > 2.5) and the inlet height a (for a > 0.55). The effects of the geometrical parameters on the cut-off diameter are depicted in Fig. 8.20. To study the effect of each parameter, the tested RBFNN model has been used by varying one parameter at a time from its minimum to maximum values of the available 98 dataset, while the other parameters 190
10
h, S
0.5 14 12 10 1 1.5 2 2.5 3 3.5
X 50 [micron]
8 6 4 2 0
Dx a b S Ht h Bc
0.2
0.4
0.6
0.8
D x, a, b, Bc
Figure 8.20: The effect of geometrical parameters on the cut-off diameter based on the Iozia and Leith model [85]. Note: The plotted curves are obtained for a test case with the following settings, Barrel diameter=0.1 m, air ow rate=0.8333l/s, air viscosity=1.0E-5 Pa s, particle density=860 kg/m3 .
are kept constant at their mean values (cf. Table 8.7). Figure 8.20 indicates the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b, the inlet height a and the total height Ht . Less effect is due to the cylinder height h and the cone tip diameter Bc . More analysis is given in Table 8.19.
The signicant geometrical parameters on the cut-off diameter (Stokes number) using the response surface methodology (RSM) Table 8.20 represents the parameters ranges selected for the seven geometrical parameters. The study was planned using BoxBehnken design, with 64 combinations. A signicant level of P < 0.05 (95% condence) was used in all tests. Analysis of variance (ANOVA) was followed by an F-test of the individual factors and interactions [53]. 191
Chapter 8. Optimization
Table 8.19: The variation of the cut-off diameter with cyclone dimensions using the RBFNN model (cf., Fig. 8.20)
Factor
Analysis The vortex nder diameter has the most signicant effect on the cut-off diameter x50 (the highest slope in Fig. 8.20). The slope is very high until Dx = 0.5 and any further increase in Dx produces a small change in x50 . In general, increasing Dx increases x50 (decreasing the collection efciency), this is one of the main reasons of the trade-off between the Euler number and the cut-off diameter objectives. This makes the optimization of cyclone geometry a multi-objective procedure. The variation of x50 with the inlet width is similar in trend and signicance to that for Dx but here the slope changes at b = 0.25. The effect of the vortex nder length and the inlet section height on the cut-off diameter is almost paralleled up to S = 1.5 and a = 0.6 afterwards they lose their signicance and become nearly constant. Increasing the barrel height slightly decreases the cut-off diameter with nearly linear relation. This trend has been reported by other researchers using CFD simulations, e.g., Elsayed and Lacor [51]. The effect of the cyclone total height is basically due to two effects the cone height and barrel height. The curve can be subdivided into four main regions. Sharp decrease in x50 up to Ht = 3.25, no valuable difference between 3.25 and 5.25, sharp increase between 5.25 and 8, and insignicant effect beyond 8. The effect of the cone-tip diameter on the cut-off diameter is quite small. First, increasing the cone-tip diameter slightly decreases the cut-off diameter up to Bc = 0.55 and any further increment increases the cut-off diameter. This trend has been reported by other researchers, e.g., Elsayed and Lacor [46, 55].
Dx
S and a
Ht
Bc
Due to the interaction between the geometrical parameters, especially between Ht with S and Dx (cf., Fig. 8.22 and Table 8.21), the obtained conclusions may not be applicable generally and the application of response surface methodology to analysis the effect of each particular parameter must.
192
Analysis of variance (ANOVA) Analysis of variance (ANOVA) showed that the resultant quadratic polynomial models adequately represented the experimental data with the coefcient of multiple determination R2 being 0.984099 (cf., Table 8.21). This indicates that the quadratic polynomial model obtained was adequate to describe the inuence of the independent variables studied [189]. Analysis of variance (ANOVA) was used to evaluate the signicance of the coefcients of the quadratic polynomial models (see Table 8.21). For any of the terms in the models, a large F-value (small P-value) would indicate a more signicant effect on the respective response variables [50, 53]. Based on the ANOVA results presented in Table 8.21, the variable with the largest effect on the Stokes number (cut-off diameter) was the linear term of vortex nder diameter, followed by the linear term of the cyclone total height, the vortex nder length and the inlet width (P < 0.05); the other three linear terms (inlet height, barrel height, and cone tip diameter) did not show a signicant effect (P > 0.05). The quadratic term of cyclone total height, vortex nder diameter and vortex nder length also had a signicant effect on the pressure drop; however, the effect of the other four quadratic terms was insignicant. Furthermore, the interaction between Dx with (Ht , S , b) and between S with (h, Ht ) also had a signicant effect on the Stokes number, while the effect of the remaining terms was insignicant. Analysis of response surfaces For visualization of the calculated factor, main effects plot, Pareto chart and response surface plots were drawn. The slope of the main effect curve is proportional to the size of the effect and the direction of the curve species a positive or negative inuence of the effect [50, 61], Fig.8.21(a). Based on the main effect plot, the most signicant factors on the Stokes number are: (1) the vortex nder diameter Dx , with a secondorder curve of direct relation. (2) the cyclone total height Ht inversely related to the Stokes number. (3) the vortex nder length S with direct relationship. (4) the inlet 193
Chapter 8. Optimization
dimensions width b and height a inversely related to the Stokes number. Whereas the other factors have an insignicant effect. The main effect plot supports the analysis given in Table 8.19, except for Ht where the strong interaction between the cyclone total height and the vortex nder length affected the trend given in Fig. 8.20. Pareto charts were used to summarize graphically and display the relative importance of each parameter with respect to the Stokes number [50]. The Pareto chart shows all the linear and second-order effects of the parameters within the model and estimates the signicance of each with respect to maximizing the Stokes number response. A Pareto chart displays a frequency histogram with the length of each bar proportional to each estimated standardized effect [30]. The vertical line on the Pareto charts judges whether each effect is statistically signicant within the generated response surface model; bars that extend beyond this line represent effects that are statistically signicant at a 95% condence level. Based on the Pareto chart (Fig. 8.21(b)) and ANOVA table (Table 8.21) there are ve signicant parameters at a 95% condence level: the vortex nder diameter Dx , the total cyclone height Ht , the vortex nder length S and the inlet dimensions a and b. Therefore, the Pareto chart is a perfect supplement to the main effect plot. To visualize the effect of the independent variables on the dependent ones, surface response of the quadratic polynomial models were generated by varying two of the independent variables within the experimental range while holding the other factors at their central values (cf., Table 8.20) [189] as shown in Fig. 8.22. Thus, Fig. 8.22(a) was generated by varying the total height Ht and the vortex nder length S while keeping the other ve factors constant. The response surface plots presented in Fig. 8.22 illustrate the strong interactions between Ht with (S and Dx ) and Dx with (S and b).
8.4.2
The genetic algorithm optimization technique has been applied to obtain the geometrical ratios for minimum pressure drop (Euler number). The objective function is the Euler number (using the trained radial basis function neural network presented in Sec. 8.3.1.1, page 163). The design variables are the seven geometrical dimensions of the cyclone separator. 194
(b) Pareto chart. A=a, B=b, C=Bc , D=Dx , E=h, F=Ht , G=S , FG=Ht *S , etc.
195
Chapter 8. Optimization
(a) Ht versus S
(b) Dx versus Ht
(c) Dx versus S
(d) b versus Dx
Figure 8.22: The response surface plots for the Stokes number. Note: the stokes number values are multiplied by 1000.
196
GA settings Table 8.22 presents the settings used to obtain the optimum design for minimum pressure using global optimization Matlab toolbox (Matlab 2010a commercial package). The evolution of the cost function for the best individual is given in Fig. 8.23. After 70 iterations (14400 function counts), the Matlab code stops generating a new population since the average change in the tness value becomes less than 1E-6. The total calculation time for this optimization problem was 102 seconds. Table 8.23 gives the optimum values for cyclone geometrical parameters for minimum pressure drop estimated by the articial neural network using the genetic algorithm optimization technique. It is clear from Table 8.23 that the new optimized design is very close to the Stairmand design in many geometrical parameters, whereas the new ratios will result in minimum pressure drop. To understand the effect of this small change in the geometrical ratios on the ow eld pattern and performance, a CFD study for the two designs is needed [53].
8
7.5
Fitness value
6.5
5.5
4.5
10
20
30
40
50
60
70
Generation
Figure 8.23: Evolution of the cost function for the best individual
8.4.2.1 Comparison between the two cyclone designs using CFD Numerical settings The air volume ow rate Qin =0.08 m3 /s for the two cyclones (inlet velocity for Stairmand design is 19 m/s and 16 m/s for the new design), air density 197
Chapter 8. Optimization
1.0 kg/m3 and dynamic viscosity of 2.11E-5 P a s. the turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [75]. Velocity inlet boundary condition is applied at inlet, outow at gas outlet and wall boundary condition at all other boundaries [53]. The nite volume method has been used to discretize the partial differential equations of the model using the SIMPLEC (Semi-Implicit Method for Pressure-Linked Equations-Consistent) method for pressure velocity coupling and QUICK scheme to interpolate the variables on the surface of the control volume. The implicit coupled solution algorithm was selected. The unsteady Reynolds stress turbulence model (RSM) was used in this study with a time step of 0.0001 s. The grid renement study using different levels of grid shows that a total number of 134759 hexahedral cells for the Stairmand cyclone and 378963 hexahedral cells for the new design are sufcient to obtain a grid independent solution, and further mesh renement yields insignicant changes in the numerical solution. The hexahedral meshes have been obtained using the GAMBIT commercial software. These simulations were performed on an 8 nodes CPU Opteron 64 Linux cluster using Fluent 6.3.26 commercial software. The geometrical values for the two cyclones are given in Table 8.24.
Figure 8.24: The contour plots for the time averaged ow variables at sections Y=0. From top to bottom: Stairmand design and the new design respectively. From left to right: the static pressure (N/m2 ), the tangential and axial velocity (m/s). Note: both cyclones have the same barrel diameter and air volume ow rate.
199
Chapter 8. Optimization
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
2000
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
2000
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
2000
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
Figure 8.25: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S1S3 (cf., Table 8.25). From top to bottom: section S1S3. From left to right: time-averaged static pressure, tangential and axial velocity respectively.
This indicates that, the new design has a lower pressure drop with respect to the Stairmand design. The velocity eld Based on the contour plots of the timeaveraged tangential velocity presented in Fig. 8.24, and the radial proles at sections S1S6 shown in Figs. 8.25 and 8.26, the following conclusions can be drawn. The tangential velocity prole at any section is composed of two regions, an inner and an outer one. In the inner region, the ow rotates approximately like a solid body (forced vortex), where the tangential velocity increases with radius. After reaching its peak the velocity decreases with radius in the outer 200
2.5
New design Stairmand design
1
New design Stairmand design
2000
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
2000
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
2500
New design Stairmand design
2.5
New design Stairmand design
1
New design Stairmand design
2000
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
Figure 8.26: The radial prole for the timeaveraged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S4S6 (cf., Table 8.25). From top to bottom: section S4S6. From left to right: time-averaged static pressure, tangential and axial velocity respectively.
part of the prole (free vortex). This prole is a so-called Rankine type vortex. The maximum tangential velocity may reach twice the average inlet velocity and occurs in the annular cylindrical part. The tangential velocity distributions for the two cyclones are approximately nearly identical in pattern and values (dimensionless), with the highest velocity occurring at 1/4 of the cyclone radius for both cyclones. This implies a nearly equal collection efciency for both cyclones, as the centrifugal force is the main driving force for particle collection in the cyclone separator. The axial velocity proles for the two cyclones are also very close, exhibiting the inverted W axial velocity prole.
201
Chapter 8. Optimization
1 Stairmand design New design
0.75
0.5
0.25
10
-1
10
10
Figure 8.27: The grade efciency curves for the two designs
The DPM results In order to calculate the cut-off diameters of the two cyclones, 104 particles were injected from the inlet surface with a velocity equals the air inlet velocity and particle density p = 860kg/m3. The grade efciency curves for the two designs are plotted in Fig. 8.27. The DPM analysis results and the pressure drops for the two cyclones are depicted in Table 8.26. Although, the difference between the two cyclone cut-off diameters is small, the saving in pressure drop is considerable (nearly 32.5% the value of Stairmand cyclone). Based on the ow pattern analysis and the DPM results, one can conclude that the cyclone collection efciency for the two cyclones is very close, with the advantage of low pressure drop in the new design. The authors want to emphasis that only small changes in the geometrical dimensions of the two designs lead to this improvement in the performance.
8.4.3
NSGA-II settings Table 8.27 presents the genetic operators and parameters for multi-objective optimization. The Euler number values have been obtained from the articial neural network trained by experimental values. The Stokes number values are obtained from Iozia and Leith model [85]. In order to investigate the effect of different geometrical and operational parameters on the Pareto front, sixteen test cases with different barrel diameter, gas ow rate and particle density have been tested, cf. Table 8.28. The sixteen test cases covers: 1) Two barrel diameters, 31 mm and 205 mm. 2) Four levels of air ow rates, 50, 60, 70 and 80 l/min. 3) Five values of particle density, 860, 202
1000, 1500, 1750 and 2000 kg/m3 . Pareto front The Pareto front (non dominated points) for the sixteen test cases are presented in Fig. 8.28(a). Figure 8.28(a) clearly demonstrate tradeoffs in objective functions Euler number and Stokes number from which an appropriate design can be compromisingly chosen by the designer [148]. All the optimum design points in the Pareto front are non-dominated and could be chosen by a designer as optimum cyclone separator [148]. The corresponding geometrical ratios of the Pareto front shown in Fig. 8.28(a) are given in Table 8.29 for test case 1 and Table 8.30 for test case 9. Three points A, B and C are indicated in Figs. 8.28(c) and 8.28(c) and Tables 8.29 and 8.30. Point A indicates the point of minimum Euler number and maximum Stokes number. Point B indicates the point of maximum Euler number and minimum Stokes number. Point C indicates an optimal point for the multiobjective optimization problem. In order to obtain the Euler number- Stokes number relationship, Fig. 8.28(b) has been drawn. It indicates a general relationship (trend) between the two dimensionless numbers irrespective to the barrel diameter, gas ow rate, particle density. A second-order polynomial has been tted between the logarithms of Euler number and Stokes number, Eq. 8.13. The obtained correlation can t the data with a coefcient of correlation R2 = 0.98643 as shown in Fig. 8.28(b). Stk50 = 100.3016(log10 (Eu)) Bubble plots for Pareto front For visual inspection of the effect of the seven geometrical parameters on the two conicting performance parameters, the bubble plots on Pareto front points have been drawn for each geometrical parameter. However, only gures for test case 1 (Fig. 8.29) and 9 (Fig. 8.30) are presented, but all other cases depict the same results (trend). Figure 8.29 indicates that: a) Decreasing the vortex nder diameter Dx decreases the Stokes number and increases the Euler number, Fig. 8.29(b). b) Generally speaking, increasing the inlet height a increases the Euler number and decreases the Stokes number. c) A similar trend is exhibited 203
2
0.9479log10 (Eu)2.5154
(8.13)
Chapter 8. Optimization
by the inlet width b but due to interaction with other geometrical and operational variables, one could see a range of bubble sizes in the region of best performance (lower values for both the Euler and Stokes numbers). d) The higher values of total cyclone height Ht will produce less Stokes number, intermediate values could produce less Euler number, smallerintermediate values could produce the optimum performance due to interaction with other variables. e) Short barrels will produce better collection efciency (low Stokes number) and higher Euler numbers. Intermediate values results in low Euler number values. Long barrels can produce the best performance. f) Short vortex nder may produce higher values of Euler numbers or higher values of Stokes number due to strong interaction with other variables. Long vortex nder can produce the optimum performance. g) Generally speaking, the variation of the cone-tip diameter Bc has no effect on the performance parameter. The above comments is restricted to the range of each geometrical variables located on the Pareto front and not for the whole range of values (cf., Fig. 8.29 for the range of each geometrical parameters).
8.4.4
Conclusions
To predict the complex non-linear relationships between the performance parameters and the geometrical dimensions, two radial basis neural networks (RBFNNs) are developed and employed to model the Euler number and Stokes number for cyclone separators. The neural networks have been trained and tested by the experimental data available in literature for Euler number (pressure drop) and Iozia and Leith model [85] for the Stokes number (cut-off diameter). The effects of the seven geometrical parameters on the Stokes number have been investigated using the trained ANN. To declare any interaction between the geometrical parameters affecting the Stokes number, the response surface methodology has been applied. The trained ANN has been used as an objective function to obtain the cyclone geometrical ratios for minimum Euler number using the genetic algorithms optimization technique. A CFD comparison between the new optimal design and the Stairmand design using the Reynolds stress turbulence model has been performed. A multi-objective optimization technique using NSGA-II technique has been applied to determine the Pareto front for the best performance cyclone separator. The following conclusions can be drawn from analysis of the obtained results: 204
205
Chapter 8. Optimization
Table 8.21: Analysis of variance and the regression coefcients of the tted quadratic equation for the Stokes numbera
Source 0 Linear 1 2 3 4 5 6 7 Quadratic 11 22 33 44 55 66 77 Interaction 12 13 14 15 16 17 23 24 25 26 27 34 35 36 37 45 46 47 56 57 67 R2 Regression coefcient -0.0470554 0.942933 -1.3178 2.10188 8.3493 -0.843633 -0.527695 1.46453 -1.17696 1.44904 -3.50314 2.33086 0.281211 0 0.243505 0.726218 0 -2.09219 0 0.121646 -0.187708 0 -4.84457 0 0.262495 -0.466053 0 0 0 0 0 -0.869946 1.44156 0 0 -0.370128 0.984099 F-Ratio P-Value
6.65 24.12 0 1145.05 0 243.98 137.75 0.21 0.18 0.37 9.38 1.49 39.06 7.33 0.03 0 1.25 0 0.22 0.09 0 5.03 0 0.78 0.39 0 0 0 0 0 38.41 16.87 0 39.06 58.83
0.0154 0.0000 1 0.0000 1 0.0000 0.0000 0.6489 0.6737 0.5476 0.0048 0.232 0.0000 0.0114 0.8558 1 0.2732 1 0.6401 0.7726 1 0.0330 1 0.3842 0.5352 1 1 1 1 1 0.0000 0.0003 1 0.0000 0.0000
a Bold numbers indicate signicant factors as identied by the analysis of variance (ANOVA) at the 95% condence level.
Table 8.22: Genetic operators and parameters for single objective optimization
Population type: Initial range: Fitness scaling: Selection operation: Elite count: Crossover fraction: Crossover operation: Mutation operation: Maximum number of generations (iterations): Population size: Double vector [0.2 0.1 0.1 0.3 2.0 0.65 0.05 ; 0.8 0.8 0.6 0.8 8.0 2.5 0.75] Rank Tournament (tournament size equals 4) 2 0.8 Intermediate crossover with the default value of 1.0 The constraint dependent default 1400 200
206
Table 8.23: The optimized cyclone separator design for minimum pressure drop
Factor Dx a b S Ht h Bc Low 0.2 0.5 0.14 0.4 3.0 1.0 0.2 High 0.75 0.75 0.4 2.0 7.0 2.0 0.4 Stairmand design 0.5 0.5 0.2 0.5 4.0 1.5 0.375 Optimum design 0.549 0.595 0.201 0.595 4.549 1.411 0.275
Table 8.24: The geometrical parameters for the two designs (D=0.205 m)
Cyclone Stairmand design New design a/D 0.5 0.595 b/D 0.2 0.201 Dx /D 0.5 0.549 Ht /D 4 4.549 h/D 1.5 1.411 S/D 0.5 0.595 Bc /D 0.375 0.275
The outlet section is above the cyclone surface by Le = 0.618D. The inlet section located at a distance Li = D from the cyclone center.
Section z`/D b
a b
S1 2.75
S2 2.5
S3 2.25
S4 2.0
S5 1.75
S6 1.5
Sections S1S5 are located in the conical section, section S6 at the cylindrical part. z` is measured from the top of the inlet section.
Stairmand design
207
Chapter 8. Optimization
Table 8.28: The diameters, air ow rates and the particle densities for the sixteen test cases Case 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 D [mm] 205 205 205 205 205 205 205 205 31 31 31 31 31 31 31 31 Q [l/min] 50 60 70 80 50 50 50 50 50 60 70 80 50 50 50 50 p [kg/m3 ] 860 860 860 860 1000 1500 1750 2000 860 860 860 860 1000 1500 1750 2000
208
Table 8.29: The seven geometrical parameters and the obtained Euler number and Stokes number for the nondominated points (Pareto-front) for test case 1 (cf. Table 8.28)
point 1 2B 3 4 5C 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36A 37 Minimum Maximum A Dx 0.306 0.306 0.618 0.326 0.360 0.670 0.585 0.306 0.306 0.305 0.303 0.476 0.312 0.308 0.592 0.320 0.598 0.309 0.430 0.306 0.306 0.326 0.308 0.680 0.307 0.307 0.514 0.310 0.559 0.399 0.516 0.413 0.306 0.307 0.646 0.692 0.558 0.303 0.692 a 0.659 0.688 0.229 0.331 0.295 0.230 0.226 0.685 0.666 0.516 0.286 0.240 0.622 0.686 0.229 0.570 0.229 0.441 0.307 0.667 0.392 0.331 0.686 0.229 0.653 0.368 0.235 0.596 0.246 0.308 0.244 0.256 0.507 0.644 0.229 0.228 0.216 0.216 0.688 b 0.385 0.398 0.213 0.240 0.253 0.213 0.217 0.299 0.361 0.271 0.318 0.217 0.277 0.379 0.213 0.273 0.213 0.266 0.231 0.389 0.277 0.209 0.348 0.213 0.296 0.338 0.217 0.275 0.221 0.229 0.241 0.280 0.352 0.333 0.219 0.213 0.230 0.209 0.398 S 0.410 0.404 0.411 0.451 0.443 0.410 0.419 0.401 0.407 0.439 0.449 0.434 0.419 0.427 0.413 0.423 0.412 0.450 0.438 0.417 0.441 0.451 0.427 0.409 0.421 0.439 0.429 0.424 0.453 0.438 0.422 0.444 0.437 0.423 0.409 0.408 0.439 0.401 0.453 Ht 6.957 6.993 6.774 6.668 6.664 6.777 6.763 6.985 6.965 6.840 6.611 6.743 6.943 6.974 6.831 6.898 6.770 6.769 6.735 6.950 6.708 6.668 6.974 6.787 6.972 6.708 6.709 6.918 6.785 6.707 6.728 6.667 6.811 6.960 6.797 6.819 6.802 6.611 6.993 h 1.779 1.779 1.885 1.885 1.910 1.859 1.901 1.789 1.782 1.868 1.930 1.901 1.909 1.781 1.895 1.909 1.893 1.881 1.898 1.779 1.882 1.900 1.807 1.857 1.816 1.895 1.901 1.899 1.898 1.891 1.904 1.927 1.848 1.807 1.863 1.855 1.922 1.779 1.930 Bc 0.387 0.317 0.495 0.444 0.459 0.496 0.492 0.425 0.398 0.428 0.462 0.471 0.429 0.459 0.494 0.434 0.494 0.435 0.465 0.367 0.443 0.471 0.459 0.495 0.420 0.430 0.479 0.429 0.455 0.460 0.486 0.472 0.460 0.401 0.497 0.498 0.486 0.317 0.498 Euler number 23.843 27.322 1.026 6.485 4.892 0.879 1.159 18.474 22.300 12.977 8.404 1.915 14.816 23.145 1.117 12.767 1.095 10.629 3.211 24.941 9.963 5.515 21.132 0.849 17.606 11.399 1.596 14.387 1.486 3.732 1.797 3.510 16.110 19.929 0.963 0.815 1.311 0.815 27.322 Stokes number x 103 0.552 0.540 2.785 0.798 0.937 3.478 2.089 0.576 0.558 0.633 0.699 1.524 0.619 0.554 2.491 0.658 2.565 0.674 1.229 0.550 0.676 0.823 0.564 3.640 0.588 0.659 1.718 0.622 1.915 1.102 1.674 1.162 0.602 0.575 3.113 3.805 1.945 0.540 3.805
indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum Stokes number. C indicates an optimal point for the multi-objective optimization problem. (cf. Fig. 8.28(c)).
209
Chapter 8. Optimization
Table 8.30: The seven geometrical parameters and the obtained Euler number and Stokes number for the nondominated points (Pareto-front) for test case 9 (cf. Table 8.28)
point 1A 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26B 27 28 29C 30 31 32 33 34 35 36 37 Minimum Maximum A Dx 0.686 0.308 0.308 0.585 0.314 0.309 0.578 0.309 0.310 0.473 0.308 0.432 0.321 0.308 0.308 0.407 0.309 0.452 0.678 0.625 0.309 0.313 0.664 0.309 0.608 0.308 0.309 0.308 0.330 0.658 0.308 0.520 0.631 0.549 0.316 0.686 0.380 0.308 0.686 a 0.236 0.655 0.692 0.235 0.330 0.486 0.237 0.272 0.538 0.238 0.375 0.238 0.327 0.606 0.643 0.255 0.333 0.236 0.236 0.237 0.596 0.325 0.237 0.365 0.235 0.692 0.436 0.427 0.290 0.245 0.362 0.251 0.236 0.242 0.270 0.236 0.239 0.235 0.692 b 0.236 0.390 0.390 0.235 0.236 0.335 0.235 0.347 0.369 0.235 0.382 0.236 0.236 0.383 0.381 0.240 0.353 0.236 0.236 0.236 0.379 0.262 0.236 0.345 0.235 0.390 0.381 0.360 0.252 0.239 0.365 0.244 0.236 0.240 0.341 0.236 0.237 0.235 0.390 S 0.478 0.423 0.423 0.418 0.431 0.426 0.421 0.424 0.429 0.478 0.423 0.445 0.431 0.424 0.424 0.470 0.424 0.448 0.473 0.449 0.425 0.430 0.476 0.425 0.438 0.423 0.424 0.424 0.465 0.477 0.424 0.470 0.449 0.430 0.425 0.478 0.477 0.418 0.478 Ht 6.910 6.902 6.902 6.925 6.837 6.869 6.684 6.897 6.882 6.923 6.901 6.785 6.839 6.896 6.895 6.899 6.882 6.793 6.893 6.800 6.894 6.847 6.911 6.876 6.921 6.902 6.899 6.886 6.890 6.910 6.883 6.906 6.796 6.909 6.891 6.910 6.890 6.684 6.925 h 1.915 1.995 1.995 1.727 1.864 1.931 1.975 1.967 1.984 1.971 1.989 1.947 1.862 1.990 1.986 1.907 1.958 1.945 1.919 1.944 1.988 1.884 1.914 1.946 1.797 1.995 1.989 1.962 1.955 1.924 1.668 1.684 1.945 1.783 1.966 1.915 1.922 1.668 1.995 Bc 0.473 0.471 0.472 0.471 0.470 0.471 0.471 0.472 0.467 0.471 0.471 0.472 0.470 0.471 0.472 0.472 0.471 0.472 0.473 0.472 0.471 0.471 0.473 0.471 0.472 0.472 0.464 0.471 0.471 0.473 0.467 0.472 0.472 0.471 0.472 0.473 0.473 0.464 0.473 Euler number 1.008 21.688 22.898 1.355 6.568 13.947 1.356 8.025 16.799 2.105 12.168 2.494 6.223 19.736 20.775 3.119 10.021 2.255 1.029 1.187 19.039 7.197 1.076 10.740 1.261 22.898 14.120 13.112 5.639 1.146 11.924 1.996 1.164 1.613 7.482 1.008 3.327 1.008 22.898 Stokes number x 103 3.743 0.558 0.551 2.436 0.750 0.621 2.008 0.697 0.597 1.481 0.634 1.300 0.777 0.571 0.565 1.165 0.665 1.394 3.615 2.891 0.578 0.730 3.404 0.656 2.691 0.551 0.616 0.626 0.818 3.290 0.648 1.671 2.964 1.823 0.726 3.743 1.071 0.551 3.743
indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum Stokes number. C indicates an optimal point for the multi-objective optimization problem. (cf. Fig. 8.28(d)).
210
4 3.5
3
Stokes number x 10
Stokes number x 10
Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Case 7 Case 8 Case 9 Case 10 Case 11 Case 12 Case 13 Case 14 Case 15 Case 16
25
0.5
10
15
20 25 30 35
Euler number
4 3.5
3
4 3.5
Stokes number x 10
Stokes number x 10
211
Chapter 8. Optimization
212
(d) Ht , range: 6.611 - 6.993 (e) h, range: 1.779 - 1.930 (f) S , range: 0.401 - 0.453
Figure 8.29: Bubble plots for different geometrical parameters for test case 1 (cf. Fig. 8.28(c) and Table 8.29).
213
(d) Ht , range: 6.684 - 6.925 (e) h, range: 1.668 - 1.995 (f) S , range: 0.418 - 0.478
Figure 8.30: Bubble plots for different geometrical parameters for test case 9 (cf. Fig. 8.28(d) and Table 8.30).
Chapter 8. Optimization
Table 8.31: The seven geometrical parameters and the obtained Euler number and Stokes number for the three points, A, B and C for the 16 test cases (cf. Table 8.28)
Point A1 B1 C1 A2 B2 C2 A3 B3 C3 A4 B2 C4 A5 B5 C5 A6 B6 C6 A7 B7 C7 A8 B8 C8 A9 B9 C9 A10 B10 C10 A11 B11 C11 A12 B12 C12 A13 B13 C13 A14 B14 C14 A15 B15 C15 A16 B16 C16 Dx 0.692 0.306 0.360 0.700 0.301 0.415 0.676 0.304 0.363 0.671 0.303 0.347 0.669 0.301 0.343 0.697 0.301 0.380 0.660 0.301 0.401 0.660 0.301 0.333 0.686 0.308 0.330 0.699 0.302 0.375 0.687 0.300 0.367 0.694 0.301 0.392 0.700 0.300 0.348 0.692 0.305 0.398 0.676 0.301 0.332 0.684 0.314 0.348 a 0.228 0.688 0.295 0.204 0.697 0.414 0.209 0.624 0.278 0.205 0.622 0.237 0.224 0.672 0.228 0.207 0.671 0.241 0.214 0.690 0.350 0.214 0.690 0.278 0.236 0.692 0.290 0.211 0.636 0.298 0.212 0.670 0.259 0.202 0.686 0.368 0.235 0.574 0.308 0.202 0.695 0.389 0.237 0.688 0.224 0.217 0.687 0.292 b 0.213 0.398 0.253 0.227 0.399 0.287 0.204 0.398 0.230 0.212 0.399 0.271 0.204 0.387 0.282 0.207 0.400 0.303 0.205 0.322 0.258 0.205 0.322 0.263 0.236 0.390 0.252 0.205 0.399 0.250 0.232 0.399 0.282 0.202 0.400 0.281 0.201 0.365 0.239 0.203 0.398 0.272 0.205 0.367 0.278 0.202 0.388 0.245 S 0.408 0.404 0.443 0.446 0.561 0.449 0.431 0.581 0.459 0.427 0.467 0.582 0.537 0.416 0.475 0.400 0.500 0.486 0.432 0.459 0.451 0.432 0.459 0.457 0.478 0.423 0.465 0.409 0.424 0.465 0.568 0.481 0.461 0.512 0.431 0.455 0.602 0.408 0.426 0.464 0.434 0.440 0.437 0.562 0.450 0.424 0.400 0.457 Ht 6.819 6.993 6.664 6.422 6.989 6.803 4.378 6.983 6.404 6.524 6.767 6.769 6.946 6.851 6.925 6.734 6.982 6.715 6.962 6.988 6.928 6.962 6.988 6.869 6.910 6.902 6.890 6.499 6.854 6.726 6.702 6.988 6.688 6.999 6.840 6.874 6.752 6.743 6.577 6.963 6.881 6.899 6.779 6.975 6.645 6.725 6.943 6.850 h 1.855 1.779 1.910 1.811 1.840 1.776 1.980 1.116 1.354 1.824 1.975 1.913 1.895 1.905 1.822 1.974 1.228 1.845 1.934 1.918 1.920 1.934 1.918 1.940 1.915 1.995 1.955 1.937 1.614 1.875 1.900 1.693 1.857 1.794 1.191 1.731 1.810 1.907 1.925 1.901 1.785 1.849 1.879 1.704 1.799 1.962 1.792 1.887 Bc 0.498 0.317 0.459 0.470 0.449 0.443 0.496 0.344 0.445 0.467 0.479 0.476 0.479 0.463 0.487 0.465 0.477 0.427 0.495 0.450 0.454 0.495 0.450 0.450 0.473 0.472 0.471 0.444 0.488 0.445 0.465 0.484 0.424 0.499 0.439 0.472 0.481 0.454 0.477 0.463 0.498 0.467 0.469 0.480 0.456 0.380 0.495 0.477 Euler number 0.815 27.322 4.892 0.820 28.222 6.052 0.868 32.296 4.791 0.819 22.621 4.873 0.900 23.579 4.647 0.706 29.367 4.557 0.810 20.739 4.735 0.810 20.739 5.582 1.008 22.898 5.639 0.741 23.886 4.619 0.955 25.971 4.801 0.736 29.948 5.822 0.901 19.174 5.047 0.722 24.604 5.614 0.890 25.765 4.912 0.823 21.743 4.969 Stokes number x 103 3.805 0.540 0.937 4.260 0.538 1.037 4.108 0.561 1.000 3.796 0.556 0.931 3.654 0.538 0.893 4.142 0.538 1.026 3.391 0.557 1.039 3.391 0.557 0.832 3.743 0.551 0.818 4.328 0.543 0.997 4.035 0.534 0.972 4.071 0.533 0.980 4.286 0.565 0.895 4.161 0.543 0.992 3.717 0.549 0.869 4.299 0.565 0.891
A indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum Stokes number. C indicates an optimal point for the multi-objective optimization problem, cf. Fig. 8.28(c).
214
Chapter 8. Optimization
simulations performed on sampling cyclone with a diameter of 31E-3 m.
Study objectives There are four objectives of this study. (1) Investigation of the effect of the four geometrical parameters on the cyclone separator performance based on CFD simulations results. (2) Study the possible interaction between the four geometrical parameters using the response surface methodology. (3) Application of the multi-objective optimization technique to obtain new geometrical ratios for minimum pressure drop and minimum cut-off diameter, followed by a comparison of the numerical simulations for the new design and the Stairmand design using the Reynolds stress turbulence model. (4) Application of the genetic algorithm technique instead of the desirability function approach using the two trained radial basis function neural networks as tness functions.
Study outline This study is performed in six stages. The rst stage is the application of the response surface methodology (Box-Behnken design) to design an experiment to study the effect of four geometrical parameters (the inlet height and width, the vortex nder diameter, and the cyclone total height) on the cyclone performance using Statgraphics statistical software. Secondly, the obtained 27 test cases (designs) have been computationally simulated using the Reynolds stress turbulence model and discrete phase modeling with the Fluent solver. In the third stage, the Euler numbers and the cut-off diameters obtained are used to t a second order polynomial (response surface) for each response (the Euler number and the cut-off diameter). The obtained polynomials have been used to study the variation of the two responses with the four geometrical parameters. Furthermore, the obtained polynomials have been used to obtain new cyclone geometrical ratios using the multi-objective optimization between the two conicting objectives (the Euler number and the cut-off diameter) using the desirability function approach. The fourth stage is a computational investigation: a numerical comparison between the new design and the Stairmand design has been performed. Fifthly, replace the response surface methodology with the articial neural networks approach and study the effect of each geometrical parameter on the cyclone separator performance. The last stage is an optimization study using the genetic algorithm instead of desirability function. 216
Design of experiment
In order to model a complex multivariate process where the responses are inuenced by several variables, the response surface statistical technique seems the best approach [50]. The steps are as follows: (1) construct the design of experiment by identifying the four tested geometrical parameters (minimum and maximum values) and also to decide upon the dependent variables (the Euler number and the cut-off diameter). Statgraphics commercial statistical software gives 27 runs to be performed (cf. Table 8.32) using the Box-Behnken design of experiment method [118]. (2) Once the runs have been conducted (using CFD simulations) and recorded data inserted in the table; Statgraphics software ts a second order polynomial to this data (one response surface per dependent variable) [30]. The secondorder polynomial (response surface) has the form [50]:
4 4
Yk = 0 +
i=1
i X i +
i=1
ii Xi2 +
i<j
ij Xi Xj
(8.14)
where 0 , i , ii , and ij are the regression coefcients for intercept, linear, quadratic and interaction terms, respectively. Xi and Xj are the independent variables, and Yk is k the response variable (k = 1 for the Euler number and k = 2 for the cut-off diameter). (3) The third step, is the analysis of the response surface plot, main effect plots, Pareto chart and interaction plots. Analysis of response surface plots Analysis of variance (ANOVA) showed that the resultant quadratic polynomial models adequately represent the used data with the coefcient of determination R2 , being 0.94284 and 0.973468 for the Euler number and cut-off diameter respectively. This indicates that the quadratic polynomial models obtained were adequate to describe the inuence of the independent variables studied [189]. Analysis of variance (ANOVA) was used to evaluate the signicance of the coefcients of the quadratic polynomial models (cf. Tables 8.33 and 8.34). For any of the terms in the models, a large F-value (small P-value) would indicate a more signicant effect on the respective response variables. Based on the ANOVA results presented in Table 8.33, the variable with the largest effect on the pressure drop (Euler number) was the linear term of vortex nder diameter (3 ), the linear terms of inlet height and width (P < 0.05) (1 and 2 respectively), whereas the linear term of cyclone total height did not show a signicant effect (P > 0.05). The quadratic 217
Chapter 8. Optimization
Table 8.32: The Box-Behnkens design matrix and the responses of the Euler number and cut-off diametera
Run No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 a X1 a/D 0.250 0.500 0.375 0.375 0.375 0.250 0.375 0.375 0.375 0.500 0.250 0.375 0.375 0.375 0.500 0.375 0.250 0.375 0.500 0.250 0.375 0.375 0.500 0.250 0.500 0.375 0.375 X2 b/D 0.2625 0.2625 0.3750 0.2625 0.2625 0.2625 0.1500 0.1500 0.2625 0.3750 0.2625 0.1500 0.1500 0.3750 0.2625 0.3750 0.2625 0.2625 0.2625 0.3750 0.2625 0.3750 0.1500 0.1500 0.2625 0.2625 0.2625 X3 Dx /D 0.50 0.75 0.25 0.50 0.75 0.50 0.50 0.50 0.25 0.50 0.75 0.50 0.75 0.50 0.50 0.50 0.25 0.25 0.25 0.50 0.50 0.75 0.50 0.50 0.50 0.50 0.75 X4 Ht /D 3 4 4 4 3 5 4 5 5 4 4 3 4 3 5 5 4 3 4 4 4 4 4 4 3 4 5 Euler number (-) 3.500 2.827 27.257 3.475 2.333 2.952 2.726 2.530 17.712 9.086 1.413 3.000 1.211 7.500 5.904 6.326 12.720 21.000 25.440 4.543 4.770 3.029 3.634 1.817 7.000 4.770 1.968 Cut-off diameterb (m) 1.546 2.541 1.158 1.683 2.444 1.364 1.353 1.284 0.956 2.163 1.939 1.455 1.826 2.081 1.787 1.836 0.860 1.084 1.127 1.651 1.683 2.610 1.513 1.155 2.025 1.683 2.156
The values of h/D = 1.5, S/D = 0.5 and Bc /D = 0.375 are identical to that of Stairmand high efciency design, so the variation in the total cyclone height is due to the variations of the cone height.
b The value of cut-off diameter depends on the gas velocity and density, particle density. In this study, D = 31mm, gas volume ow rate Qin = 50 l/min, gas density = 1.0, gas viscosity = 2.11E 5 Pa s and particle density p = 860 kg/m3 .
term of vortex nder diameter also had a signicant effect on the pressure drop; however, the effect of the other three quadratic terms was insignicant. Furthermore, the interaction between the inlet dimensions and vortex nder diameter also had a signicant effect on the pressure drop (13 and 23 ), whereas the effect of the remaining terms was insignicant. Table 8.34 conrms the signicant effect of all linear terms of the inlet dimensions, vortex nder diameter and total cyclone height on the cut-off diameter. Moreover, the interaction between the inlet width and the vortex nder diameter also had a signicant effect on the cut-off diameter, whereas the effect of the remaining terms was insignicant. For the visualization of the results of the analysis, main effects plot, Pareto chart and response surface plots were drawn. The slope of the main effect curve is proportional to the size of the effect, and the direction of the curve species a positive or negative inuence of the effect [50, 61] (Fig.8.31(a)). Based on the main effect plot and the Pareto chart shown in Fig. 8.31, the most signicant factors on the Euler number are: (1) the vortex nder diameter Dx , with a non-linear relation with a wide range where it is inversely proportional to Euler number and a narrow range with a direct 218
Regression coefcient 1.88508 41.3522 134.36 -49.1776 -4.14082 16.9991 -44.8511 87.9024 0.368311 48.4622 -90.448 -1.096 -185.69 -1.56444 2.923 0.94284
F-Ratio
P-Value
14.21 33.78 209.03 0.94 0.1 0.5 44.97 0.2 0.41 7.01 0.02 28.5 0.03 0.47
0.0012 0.0000 0.0000 0.3430 0.7509 0.4878 0.0000 0.6602 0.5306 0.0155 0.8992 0.0000 0.8708 0.5017
Bold numbers indicate signicant factors as identied by the analysis of variance (ANOVA) at the 95% condence level.
proportional relation. (2) the inlet width b and the inlet height a almost linearly related to the Euler number, (3) the cyclone total height Ht has an insignicant effect. Pareto charts were used to summarize graphically and display the relative importance of each parameter with respect to the Euler number [50]. The Pareto chart shows all the linear and secondorder effects of the parameters within the model and estimates the significance of each with respect to maximizing the Euler number response. A Pareto chart displays a frequency histogram with the length of each bar proportional to each estimated standardized effect [30]. The vertical line on the Pareto charts judges, whether each effect is statistically signicant within the generated response surface model; bars that extend beyond this line represent effects that are statistically signicant at a 95% condence level. Based on the Pareto chart (Fig. 8.31(b)) there are six signicant parameters at a 95% condence level: the vortex nder diameter Dx ; the 2 inlet width b; the inlet height a and the combinations a Dx , b Dx and Dx . Therefore, the Pareto chart is a perfect supplement to the main effects plot. To visualize the effect of the independent variables on the dependent ones,response surfaces of the quadratic polynomial models were generated by varying two of the independent variables within the experimental range whereas holding the other factors at their central values [189] as shown in Fig. 8.31. Thus, Fig. 8.31(e) was generated by varying the inlet height a and the inlet width b whereas keeping the other factors xed at their central values. The trend of the curve is linear, with a more signicant effect for the inlet width b, and a weak interaction between the inlet height a
219
Chapter 8. Optimization
Table 8.34: diameter
Variable 0 Linear 1 2 3 4 Quadratic 11 22 33 44 Interaction 12 13 14 23 24 34 R2
67.21 89.76 524.79 15.14 0.82 0.06 0.25 0.05 0.59 2.77 0.08 18.61 0.14 0.63
0.0000 0.0000 0.0000 0.0009 0.3763 0.8152 0.6237 0.8308 0.4531 0.1116 0.7837 0.0003 0.7170 0.4360
Bold numbers indicate signicant factors as identied by the analysis of variance (ANOVA) at the 95% condence level.
and width b. The response surface plots of Figs. 8.31(c), 8.31(d) and 8.31(f) show that there are strong interactions between the vortex nder diameter Dx and the inlet dimensions a and b. From the analysis of the design of experiment for the cut-off diameter, the effect of variation of the vortex nder diameter Dx on the cut-off diameter is opposite to that on the Euler number (cf. Fig. 8.32(a)). The Pareto chart given in Fig. 8.32(b) indicates the signicance effect of the cyclone total height Ht on the cut-off diameter, but its effect is minor in comparison with the three other factors. The signicant interaction exists only between the vortex nder diameter and the inlet dimensions.
8.5.2
The desirability function approach From the previous analysis, it is observed that the optimal values for the geometrical parameters that minimize the pressure drop are different from the values that minimize the cut-off diameter (cf. Figs. 8.31 and 8.32). As a result, a multi-objective optimization procedure is needed. The uti220
lization of desirability function proposed by Harrington [71] is the most popular and strongly suggested method for multiple response optimization problems [20] to convert the problem into single-objective. The Statgraphics statistical package uses this desirability function approach for optimization of multiple response problems. The desirability function is rst dened for each response. The desirability function d(y ) expresses the desirability of a response value equal to y on a scale of 0 (if the response value is in an unacceptable range) to 1 (for the optimum value), for minimization of response variable. The procedure will then nd the settings of the experimental factors that maximize a combined desirability function, which is a function that expresses the desirability of a solution involving m, where m here equals 2 (one for the Euler number and the other for the
221
Chapter 8. Optimization
(c) Dx versus a
(d) Dx versus b
(e) a versus b
(f) Dx versus Ht
Figure 8.31: (continued) Analysis of design of experiment for the Euler number. Note: In Figs 8.31(c) - 8.31(f), for each plot of two independent variable, all other variables are hold at their central values.
222
Im 1 I2 D = dI 1 d2 . . . dm
1/(
m j =1
Ij )
(8.15)
where dj is the calculated desirability of the jth response and Ij is an impact coefcient that ranges between 1 and 5 [131]. It represents the importance relative of each response over the other responses [88]. Where the default value is 3. In this study, more importance is given to the Euler number (I1 = 5 and I2 = 3). When a response is to be minimized, the 223
Chapter 8. Optimization
(c) Dx versus a
(d) Dx versus b
(e) a versus b
(f) Dx versus Ht
Figure 8.32: (continued) Analysis of design of experiment for the cut-off diameter. Note: In Figs 8.32(c) - 8.32(f), for each plot of two independent variable, all other variables are hold at their central values.
224
(8.16)
where lowj and highj are the minimum and maximum values of jth response. In this study, the minimum and maximum values obtained in the data sheet have been used (cf. Table 8.32). The desirability plots are given in Fig. 8.33. For more information about statistical model used for multiple response optimization refer to Ref. [131]. Table 8.35 presents the optimum values of geometry parameters that minimize the values of the Euler number and the cut-off diameter, which gives optimum desirability D = 0.83 to minimize the Euler number and cut-off diameter. It is clear from Table 8.35 that the new optimized design differs from the Stairmand design in many geometrical parameters, whereas the new ratios will result in minimum pressure drop and minimum cut-off diameter. The ratio of inlet width to height b/a = 0.589 lays in the optimal cyclone lies in the recommended range of b/a from 0.5 to 0.7 proposed by Elsayed and Lacor [52]. To understand the effect of these changes in the geometrical ratios on the ow eld pattern and performance, a CFD study for the two designs is needed. 8.5.2.1 CFD comparison between the Stairmand and optimal design The Fluent solver has many turbulence models available for simulating turbulent ow. It is generally recognized that only the Reynolds stress model (RSM) and large eddy simulation (LES) can capture the main features of the highly complicated swirling ow in cyclone separators [11, 23, 50, 52, 6264, 74, 91, 159, 186, 198]. The Reynolds stress turbulence model has been used in this study to reveal the turbulent ow in the two cyclone separators. For the detailed governing equation for both the Reynolds averaged Navier-Stokes equation (RANS) and the discrete phase modeling 225
Chapter 8. Optimization
(a) Dx versus a
(b) Dx versus b
(c) a versus b
(d) Dx versus Ht
Figure 8.33: The desirability plots. Note: for each plot of two independent variable, all other variables are hold at their optimal values.
226
(DPM) the reader can refer to Elsayed and Lacor [52]. The geometrical values are given in Table 8.36 for the two cyclones (cf. Fig. 8.34). Numerical settings The air volume ow rate Qin =50 l/min for the two cyclones, air density is 1.0 kg/m3 and dynamic viscosity 2.11E-5 P a s. The turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [75]. A velocity inlet boundary condition is applied at inlet, outow at gas outlet and wall boundary conditions at all other boundaries. The nite volume method has been used to discretize the partial differential equations of the model using the SIMPLEC (Semi-Implicit Method for Pressure-Linked Equations-Consistent) method for pressure velocity coupling and QUICK scheme to interpolate the variables on the surface of the control volume. The implicit coupled solution algorithm was selected. The unsteady Reynolds stress turbulence model (RSM) was used in this study with a time step of 0.0001s.
Figure 8.34: The cyclone geometry and the surface mesh for the new design.
Table 8.36: The values of geometrical parameters for the two designs (D=31E-3 m)
Cyclone Stairmand design New design a/D 0.5 0.256 b/D 0.2 0.151 Dx /D 0.5 0.415 Ht /D 4 4.56 h/D 1.5 1.5 S/D 0.5 0.5 Bc /D 0.375 0.375 Li /D 1.0 1.0 Le /D 0.5 0.5
227
Chapter 8. Optimization
Grid independency study The grid independence study has been performed for the tested cyclones. Three levels of grid for each cyclone have been tested, to be sure that the obtained results are grid independent. The hexahedral computational grids were generated using Gambit grid generator and the simulations were performed using Fluent 6.3.26 commercial nite volume solver on a 8 nodes CPU Opteron 64 Linux cluster. To evaluate accuratelyto estimate accurately the numerical uncertainties in the computational results, the concept of grid convergence index (GCI) was adopted using three grid levels per cyclone. Grid convergence index (GCI) Table 8.37 presents the grid convergency calculations using GCI method for three grid levels for each cyclone. The following conclusions have been obtained from the GCI analysis [55]: The results are in the asymptotic range, because the obtained values for are close to unity. The ratio R is less than unity this means monotonic convergence [2]. There is a reduction in the GCI value for the successive grid renements ne (GCIne 12 < GCI23 ) for the two variables (Eu and x50 ). This indicates that the dependency of the numerical results on the cell size has been reduced. Moreover, a grid independent solution has been achieved. Further renement of the grid will not give much change in the simulation results. For the two variables (Eu and x50 ), the extrapolated value is only slightly lower than the nest grid solution. Therefore, the solution has converged with the renement from the coarser grid to the ner grid [2]. Figure 8.35 presents a qualitative proof that the obtained results are in the asymptotic range. The value of i,i+1 represent the relative change in each value from coarse to medium and from medium to ne mesh. For example, 1,2 = 0.0256 for the Euler number in the new design means the percentage change in the Euler number when the mesh becomes 986748 cells instead of 717353 cells equals 2.56%. This means no need to use the ne mesh and the usage of the medium mesh of 717353 cells is sufcient. Another example, 1,2 = 0.0052 for the cut-off diameter in the Stairmand design means the percentage change in the cutoff diameter when the mesh becomes 848783 cells instead of 622253 228
cells equals 0.52%. This means no need to use the ne mesh and the usage of the medium mesh of 622253 cells is sufcient. In summary, the grid renement study shows that a total number of about 622253 hexahedral cells for Stairmand cyclone and 717353 hexahedral cells for the new design are sufcient to obtain a grid-independent solution, and further mesh renement yields only small, insignicant changes in the numerical solution.
Table 8.37: Grid convergency calculations using GCI method using three grid levels for each cyclone.
Cut-off diameter Euler number Cut-off diameter Euler number i 0c 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 Ni 986748 717353 362679 986748 717353 362679 848783 622253 356181 848783 622253 356181 fi 2.5195 2.5380 2.6030 1.2553 4.6170 0.6621 0.6710 1.1121 0.6900 1.2553 0.9820 5.4310 5.4860 1.1090 5.6060 1.2044 6.8330 1.7187 1.7210 1.1090 1.7300 1.2044 1.9150 0.1850 0.1069 0.8173 0.0090 0.0052 0.1678 0.0489 1.0052 1.2270 0.2189 3.9010 0.1200 0.0219 1.2521 0.0999 1.0219 0.2920 0.4232 5.0609 0.0190 0.0283 1.6647 0.0669 1.0283 2.0140 0.7737 4.0099 ri,i+1 ei,i+1 i,i+1 GCIi,i+1 % Ra b
1.1121
0.0650
0.0256
Stairmand design
New design
a R= / . 12 23 b = rp GCI 12 /GCI23 . 12 c The value at zero grid space (h 0). i=1, 2 and 3 denote the calculations at the ne, medium and coarse mesh respectively.
Chapter 8. Optimization
6 1.5 5
New design (Eu) New design (X50) Stairmand design (Eu) Stairmand design (X50)
1 3
1E-06
2E-06
(h --> 0)
-1
Figure 8.35: Qualitative representation of the grid independency study. The Euler number and the cut-off diameter for the two cyclones using the three grid levels. N 1 is the reciprocal of the number of cells, h 0 means the value at zero grid size (cf. Table 8.37). To obtain a smooth curve; the spline curve tting has been applied in Tecplot post-processing software.
same ow pattern, but the Euler number of the Stairmand design is nearly twice that of the new design. The dimensionless static pressure distribution presented in Fig. 8.37 for the two cyclones indicates that the highest dimensionless static pressure for the Stairmand design is more than twice that of the new design at all sections whereas the central value is almost the same for the two cyclones. This indicates that, the new design has a lower dimensionless pressure drop than the Stairmand design. However, these results are obtained at different inlet velocity for the two cyclones (to have the same air ow rate). The same Euler number values would be obtained if the two cyclones work at the same inlet velocity because the Euler number is not a function of ow velocity if the Reynolds number is higher than 2E4 [50]. The tangential velocity prole is composed of two regions. In the inner region, the ow rotates approximately like a solid body (forced vortex), where the tangential velocity increases with radius. After reaching its peak the velocity decreases with radius in the outer part of the prole (free vortex). The tangential velocity distributions for the two cyclones are nearly identical in pattern (Rankine prole). The inner part of the tangential velocity distribution of the two cyclones is very similar. The outer part for the new design is atter in comparison with the Stairmand cyclone. This implies that there is more space in the optimal cyclone where the particles are subjected to high centrifugal force. Whereas, the maximum dimensionless tangential velocity for Stairmand cyclone is higher than that for the 230
Cut-off diameter
Euler number
new design, the cyclone performance is not only affected by the maximum tangential velocity but also with the separation space (the new design is longer than the Stairmand design) The axial velocity proles for the two cyclones are different in values and shape. Stairmand cyclone exhibit the inverted W axial velocity prole. The new design exhibit the inverted W axial velocity prole away from the inlet section and the inverted V elsewhere. This is due to the very high swirl exist at the inlet section in case of the new design. Performance parameters To calculate the cut-off diameters of the two cyclones, 104 particles were injected from the inlet surface with a particles density p = 860kg/m3 and the maximum number of time steps for each injection was 2E9 steps. The DPM analysis results and the Euler number for the two cyclones are given in Table 8.38. Table 8.38 introduces also a comparison between the CFD results and four different mathematical models viz., the Ramachandran model [139], the Muschelknautz method of modeling (MM) [50, 116, 174], the Iozia model [85], the Ritema model [142](cf. Hoffmann and Stein[77] for more details about these mathematical models). However, no-good matching between the two approaches (CFD and mathematical models) is obtained, they agree in the trend of superior performance of the new design. The Euler number and cut-off diameter for the new design is approximately half that of Stairmand design.
Chapter 8. Optimization
New design Stairmand design
Figure 8.36: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From top to bottom: the dimensionless static pressure (divided by the dynamic pressure at inlet), the dimensionless tangential velocity, axial velocity. Note: both cyclones have the same air volume ow rate.
232
4.5
1.2
0.8
0.6
0.4
0.2
0 -1
-0.5
0.5
Figure 8.37: The radial proles of the time averaged static pressure and tangential velocity at z/D=1.5.
Eu = f1 x50 = f2
a b D x Ht , , , D D D D a b D x Ht , , , D D D D
(8.17)
Due to the large difference in the order of magnitude of the value (cf. Table 8.35), the available dataset is transformed into -1 to 1 interval using the Matlab intrinsic function; mapminmax in order to avoid solution divergence [53, 197]. The RBFNN calculations have been performed using the neural network toolbox available from Matlab commercial software 2010a. The cut-off diameter x50 for cyclone separator is always given in units of m. Another way to represent x50 is using a dimensionless number; Stokes number Stk50 = p x2 50 Vin /(18D ) [37]. It is the ratio between the particle relaxation time; p x2 50 /(18) and the gas ow integral time scale; D/Vin . Fitting the RBFNNs The congured RBFNNs predictions versus the CFD data for the Euler number and cut-off diameter are shown in Fig. 8.38. It can be seen that 233
Chapter 8. Optimization
the RBFNN models are able to attain the high training accuracy. The training mean square errors are zeros (i.e., identical matching between the input and output, the reason behind that may refer to the consistency between the used data in the simulation using the design of experiment, which is not the case for the study of Elsayed and Lacor [53] using experimental data set), Fig. 8.38. This indicates that, in comparison with traditional models of curve tting, the models based on an articial intelligence algorithm have a superior capability of nonlinear tting. Especially, the RBFNN has its unique and optimal approximation characteristics in learning process [53, 197]. Figure 8.38 illustrates the agreement between the RBFNNs input and output. The obtained relation is a typical linear relation with a coefcient of correlation close to 1 (R > 0.999). The agreement between the input and output of the RBFNN is also clear from the value of the mean squared error E 2 . Consequently, the trained neural networks predict very well both the Euler number and cut-off diameter values and can be used in cyclone design and performance estimation. Figure 8.38 present different performance indicators as a validation of the proposed models for the trained data.
26 24 22 20
2.5
18 16 14 12 10 8 6 4 2
1.5
10
15
20
25
1.5
2.5
Figure 8.38: Linear regression of the RBFNNs for the Euler number and the cut-off diameter.
234
Ht
3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5
2.2 15
Euler number
10
Dx a b Ht
X 50 [micron]
1.8
1.6
1.4
Dx a b Ht
1.2
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.2
0.4
0.6
D x, a, b
D x, a, b
The effect of the four geometrical parameters on the cyclone performance based on RBFNNs The effects of the geometrical parameters on both the Euler number and the cut-off diameter are depicted in Figs. 8.39(a) and 8.39(b). To study the effect of each parameter, the tested RBFNNs models have been used by varying one parameter at a time from its minimum to maximum values of the available CFD dataset, whereas the other parameters are kept constant at their mean values (cf. Table 8.35). Figures 8.39(a) and 8.39(b) indicate the signicant effect of the vortex nder diameter Dx , the inlet width b, the inlet height a. Less effect is assigned to the total cyclone height Ht . More analysis is given in Tables 8.39 and 8.40.
Chapter 8. Optimization
Table 8.39: The variation of the Euler number with cyclone dimensions using the RBFNN model (cf., Fig. 8.39(a)) Factor Analysis The most signicant effect is that of the vortex nder diameter Dx with inverse relationship when increasing Dx up to 0.65 after which the relation becomes direct, Fig. 8.39(a). This can be explained as follows: Although the pressure loss in the vortex nder decreases with increasing the vortex nder diameter like the case of viscous ow in a pipe, the Euler number in the cyclone body instead will increase due to the decrease of the ow area just after the ow entrance from the inlet region (the annular space between the barrel and the vortex nder). This analysis indicates the large contribution of the pressure loss in the vortex nder to the total Euler number (the pressure loss at the entrance, the pressure loss in the cyclone body, and the pressure loss in the vortex nder [50]). The inlet height a and width b is almost linearly related to the Euler number. But, why the relation is direct (at the same ow rate increasing the inlet dimensions decreases the inlet velocity. Consequently,the loss in the vortex nder will decrease)? The reason is the increase in the pressure drop at the inlet section, due to deviation of the inlet ow when it mixes with the swirling ow. The effect of changing the total cyclone height Ht on the Euler number is very small.
Dx
a and b
Ht
Table 8.41 presents the settings used to obtain the optimum design for minimum pressure using global optimization Matlab toolbox in Matlab 2010a commercial package. Table 8.42 gives the optimum values for cyclone geometrical parameters for minimum pressure drop estimated by the articial neural network using the genetic algorithm optimization technique. It is clear from Table 8.42 that the new optimized design is very close to the Stairmand design in many geometrical parameters, whereas the new ratios will result in the minimum pressure drop.
Optimal cyclone design for best performance using NSGA-II In case of cyclone separator geometry optimization for minimum Euler number and minimum cut-off diameter, the objectives are conicting with each other. There is no best solution for which all objectives are optimal simultaneously [181]. The increase of one objective will lead to the de236
Dx
a and b
Ht
crease of the other objective. Then, there should be a set of solutions, the so-called Pareto optimal set or Pareto front, in which one solution cannot be dominated by any other member of this set [56]. Recently, a number of multi-objective genetic algorithms (MOGAs) based on the Pareto optimal concept have been proposed. The well known nondominated sorting genetic algorithm II (NSGA-II) proposed by Deb et al. [34] is one of the most widely used MOGAs since it provides excellent results as compared with other multi-objective genetic algorithms proposed [25]. Table 8.43 presents the genetic operators and parameters for multi-objective optimization. The Euler number and the Stokes number values have been obtained from the articial neural network trained by the CFD data set. The Pareto front (non dominated points) is presented in Fig. 8.40(a) and Table 8.44. Figure 8.40(a) clearly demonstrate tradeoffs in objective functions (Euler number and Stokes number). All the optimum design points in the Pareto front are non-dominated and could be chosen by a designer as optimum cyclone separator [148]. This set of designs makes the Pareto 237
Chapter 8. Optimization
Table 8.41: Genetic operators and parameters for single objective optimization
Population type Initial range Fitness scaling Selection operation: Elite count Crossover fraction Crossover operation Mutation operation Maximum number of generations: Population size Double vector [0.25 0.15 0.25 3 ; 0.5 0.375 0.75 5 ] for a, b, Dx and Ht respectively Rank Tournament (tournament size equals 4) 2 0.8 Intermediate crossover with the default value of 1.0 The constraint dependent default 800 200
front approach more preferred than the desirability function approach which gives only one design point. Three points A, B and C are indicated in Fig. 8.40(a). Point A indicates the point of minimum Euler number (maximum Stokes number). Point B indicates the point of maximum Euler number (minimum Stokes number). Point C indicates an optimal point for the multi-objective optimization problem. In order to obtain the Euler number-Stokes number relationship, Fig. 8.40(b) has been drawn. It indicates a general relationship (trend) between the two dimensionless numbers irrespective to the four geometrical parameters values. A second-order polynomial has been tted between the logarithms of Euler number and Stokes number, Eq. 8.18. The obtained correlation can t the data with a coefcient of determination R2 = 0.99613 as shown in Fig. 8.40(b). Elsayed and Lacor [56] presented another correlation between the Euler number and the Stokes number, Eq. 8.19. Equation 8.19 presents a good matching for only high values of Euler numbers. For smaller values of Euler number, there is underestimation of the Stokes number. The difference between the two correlations can be referred to two reasons. Firstly, the new correlation is based on the CFD simulations data for both Euler number and Stokes number, whereas Elsayed and Lacor
Table 8.42: The optimized cyclone separator design for minimum pressure drop using genetic algorithm Factor a b Dx Ht Euler number Cut-off diameter Low 0.25 0.15 0.25 3.0 High 0.5 0.375 0.75 5.0 Stairmand design 0.5 0.2 0.5 4.0 5.606 1.706 Optimum design 0.492 0.158 0.617 4.535 2.369 1.704
The values for the Euler number and the cut-off diameter for the Stairmand design have been obtained from CFD simulations, cf. Table 8.38. The values for the Euler number and the cut-off diameter for the new optimal design have been obtained from the trained RBFNN.
238
correlation was obtained from experimental data for Euler number values and Iozia and Leith model for the Stokes number values. Secondly, the new correlation is limited to only four geometrical parameters. The other three factors are xed, h = 1.5, S = 0.5 and Bc = 0.375. Consequently, the new correlation (Eq. 8.18) is valid only for these values. Stk50 = 100.3533(log10 (Eu)) Stk50 = 100.3016(log10 (Eu))
2
1.1645log10 (Eu)2.3198
(8.18) (8.19)
0.9479log10 (Eu)2.5154
8.5.5 Conclusions
CFD simulations data have been used to understand the effect of four geometrical parameters on the cyclone performance and to optimize the cyclone geometry. Two meta-models have been used viz., the response surface and the radial basis function neural network approaches. Two optimization techniques have been applied, the desirability function with Nelder-Mead technique and the non-sorted dominated genetic algorithm NSGA-II. The response surface methodology has been used to t two secondorder polynomials to the Euler number and cut-off diameter obtained from CFD simulations. The analysis of variance of the Euler number indicates a strong interaction between Dx with (a, b) and between Dx with b only for the cut-off diameter. The bi-objective functions have been converted to single-objective function using the desirability function approach. A new optimal design has been obtained using the Nelder-Mead technique available in Statgraphics commercial software. The ratio of inlet width to height b/a = 0.589 lays in the optimal cyclone lies in the recommended range of b/a from 0.5 to 0.7 proposed by Elsayed and Lacor [52]. The new design and the Stairmand design have been computationally compared 239
Chapter 8. Optimization
Stokes number x 10
4 3
2 1
C 5 10 15 20 Euler number 25 30
B 35
10 8 6
Stokes number x 10 3
Euler number
(b) Pareto front in log scale. Note: The correlation (Eq. 8.19) given by Elsayed and Lacor [45] are based on seven geometrical parameters.
Figure 8.40: Pareto front plots obtained from the variation of four geometrical ratios.
240
A indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum Stokes number. C indicates an optimal point for the multi-objective optimization problem. (cf. Fig. 8.40(a)).
to get a clear vision for the differences between the ow eld pattern and performance in the two designs. The CFD simulations results and four mathematical models conrmed the better performance of the new design in comparison with the Stairmand design. The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone performance better than the response surface methodology. The used RBFNN presented zero mean squared error and almost unity coefcient of determination. The analysis indicates the signicant effect of the vortex nder diameter Dx and the inlet dimensions a and b on the cyclone performance. Moreover, the range of high inuence is given for each geometrical parameters using the trained RBFNNs. The trained RBFNN for the Euler number has been used to get a new optimized cyclone for minimum pressure drop (Euler number) 241
Chapter 8. Optimization
using the genetic algorithm optimization technique. The new cyclone design is very close to the Stairmand high efciency design in the geometrical parameter ratio, and superior in low pressure drop at nearly the same cut-off diameter. But, the optimal design obtained from the desirability function results in a better collection efciency (smaller cut-off diameter) as is clear from Table 8.38 because there the cut-off diameter is included in the desirability function, whereas, the obtained result are for single objective (Euler number). The two trained RBFNNs have been used in a multi-objective optimization process using NSGA-II technique. The Pareto front is presented for the designer with a wide choice for selection. A second-order polynomial has been tted between the logarithms of Euler number and Stokes number to obtain a general formula, 2 Stk50 = 100.3533(log10 (Eu)) 1.1645 log10 (Eu)2.3198 with a coefcient of determination R2 = 0.98643. This formula can be used to obtain the Stokes number if the Euler number is known at h = 1.5, S = 0.5 and Bc = 0.375.
242
9.1.2
The vortex nder dimensions: The maximum tangential velocity in the cyclone decreases with increasing the vortex nder diameter. A negligible change is noticed with increasing the vortex nder length. Increasing the vortex nder length, makes a small change in both the static pressure, axial and tangential velocity proles. However, decreasing the vortex nder diameter gradually changes the axial velocity prole from the inverted W to the inverted V class. Decreasing the cyclone vortex nder diameter, increases the maximum tangential velocity. The Euler number (dimensionless pressure drop) decreases with increasing the vortex nder diameter. Increasing the vortex nder length slightly increases the Euler number. The Stokes number increases with increasing the vortex nder diameter, because the centrifugal force affecting particles attenuates when the swirl intensity (maximum tangential velocity) decreases. The Stokes number slightly increases as the vortex nder length is increased. The inlet dimensions: The maximum tangential velocity in the cyclone decreases with increasing both the cyclone inlet width and height. Increasing the cyclone inlet width or height decreases the pressure drop at the cost of increasing the cut-off diameter. The effect of changing the inlet width on the cut-off diameter is more signicant in comparison with that of the inlet height. The optimum ratio of the inlet width to the inlet height is from 0.5 to 0.7. The cyclone heights: The maximum tangential velocity in the cyclone decreases with increasing the cyclone (barrel or cone) height. Increasing the barrel height, makes a small change in the axial veloc244
9.1. Conclusions
ity. Increasing the cyclone barrel height decreases the pressure drop and the cut-off diameter. The changes in the performance beyond h/D = 1.8 are small. Increasing the cone height makes a considerable change in the axial velocity. Both the pressure drop and the cutoff diameter decrease with increasing the cyclone cone height. The performance improvement stops after hc /D = 4.0 (Ht /D = 5.5). The effect of changing the barrel height is less signicant on the performance and the ow pattern in comparison with the effect of the cone height.
9.1.3 Optimization
Several new optimized cyclone geometrical ratios have been obtained and presented in this thesis. All the new geometrical ratios result in better performance than the Stairmand design. For example, The new cyclone design ratios obtained using MM model and Nelder-Mead technique for minimum pressure drop, are very close to the Stairmand high efciency design in the geometrical parameter ratios, and superior for low pressure drop at nearly the same cut-off diameter. The new cyclone design results in nearly one-half the pressure drop obtained by the Stairmand design at the same volume ow rate. Articial neural networks In order to accurately predict the complexly non linear relationships between pressure drop and geometrical dimensions, a radial basis neural network (RBFNN) is developed and employed to model the pressure drop for cyclone separators. The neural network has been trained and tested by the experimental data available in literatures. The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone performance. Four mathematical models (Muschelknautz method MM, Stairmand, Ramachandran and Shepherd and Lapple) have been tested against the experimental values. The residual error of MM model is the lowest. Moreover, one can multiply the calculated value using MM by 1.5 to get the experimental value (as a rough approximation). The response surface methodology has been used to t a second-order polynomial to the RBFNN. The second-order polynomial has been used to get a new optimized cyclone for minimum pressure drop using the Nelder-Mead technique. 245
9.1.4
Multi-objective optimization
Two multi-objective optimization approaches have been applied in this thesis, namely the desirability function and the non-dominated sorted genetic algorithm (NSGA-II). Moreover, two sources of data have been used, namely: (1) The experimental data for the pressure drop (Euler number) and the Iozia and Leith model for the cut-off diameter (Stokes number) for the optimization of the seven geometrical parameters. (2) CFD simulations for the optimization of only four geometrical parameters. Seven geometrical parameters Two radial basis neural networks (RBFNNs) are developed and employed to model the Euler number and the Stokes number for cyclone separators. The neural networks have been trained and tested by the experimental data available in literature for Euler number (pressure drop) and the Iozia and Leith model [85] for the Stokes number (cut-off diameter). The effects of the seven geometrical parameters on the Stokes number have been investigated using the trained ANN. To declare any interaction between the geometrical parameters affecting the Stokes number, the response surface methodology has been applied. The trained ANN has been used as an objective function to obtain the cyclone geometrical ratios for minimum Euler number using the genetic algorithms optimization technique. A CFD comparison between the new optimal design and the Stairmand design has been performed. A multi-objective optimization technique using NSGAII technique has been applied to determine the Pareto front for the best performance cyclone separator. The trained RBFNN has been used to get a new optimized cyclone for minimum pressure drop (Euler number) using the genetic algorithm optimization technique. The new cyclone design results in nearly 68% of the pressure drop obtained by the Stairmand design at the same volume ow rate. This 246
9.1. Conclusions
conrms that the obtained design using the genetic algorithm is better than that obtained using Nelder-Mead technique which results in 75% of the Stairmand pressure drop [53]. The two trained RBFNNs have been used in a multi-objective optimization process using NSGA-II technique. Sixteen test cases with different barrel diameter, gas ow rate and particle density have been tested. The Pareto fronts for these test cases are very close. A second-order polynomial has been tted between the logarithms of the Euler number and the Stokes number to obtain a general for2 mula, Stk50 = 100.3016(log10 (Eu)) 0.9479 log10 (Eu)2.5154 . This formula can be used to obtain the Stokes number if the Euler number is known. Four geometrical parameters CFD simulations data have been used to understand the effect of four geometrical parameters on the cyclone performance and to optimize the cyclone geometry. Two meta-models have been used viz., the response surface and the radial basis function neural network approaches. Two optimization techniques have been applied, the desirability function with Nelder-Mead technique and the non-sorted dominated genetic algorithm NSGA-II. The response surface methodology has been used to t two secondorder polynomials to the Euler number and the cut-off diameter obtained from CFD simulations. The analysis of variance of the Euler number indicates a strong interaction between Dx with (a, b) and between Dx with b only for the cut-off diameter. The bi-objective functions have been converted to single-objective function using the desirability function approach. A new optimal design has been obtained using the Nelder-Mead technique. The ratio of inlet width to height b/a = 0.589 i.e., in the optimal cyclone lies in the recommended range of b/a from 0.5 to 0.7 proposed by Elsayed and Lacor [52]. The new design and the Stairmand design have been computationally compared to get a clear vision for the differences between the ow eld pattern and performance in the two designs. The CFD simulations results conrmed the superior performance of the new design in comparison with the Stairmand design. The result demonstrates once more that articial neural networks can offer an alternative and powerful approach to model the cyclone performance better than the response surface methodology. The used 247
249
250
where Rin is the radial position of the center of the inlet (Fig. A.2), For a b slot inlet Rin = R 2 , where b is the inlet width and R = D/2 is the cyclone radius. Hoffmann and Stein [77] gave an algebraic relations for , the simplest of which is b R
0.5
= 1 0 .4
(A.2)
Rin So vw can be calculated as vw = vin R , as the inlet velocity is given or Q calculated from vin = a b where Q is the gas volume ow rate, a is the inlet height. To get from vw to the tangential velocity at CS vCS , Barth gave the following relation R vw ( R ) x
vCS =
1+
HCS Rf vw Q
(A.3)
Introducing Eq. (A.1), one obtains vCS = Rin Rx vx a b + HCS f Rin (A.4)
x Where Rx is the vortex nder radius Rx = D 2 , vx is the mean axial velocity Q in the vortex nder vx = R2 , and HCS is the height of the control surface x extending from the bottom of the vortex nder to the cyclone bottom.
253
Where co is the mass ratio of dust feeding the cyclone to the gas ow rate (dimensionless).
Pbody =
H S 0.5Dx f
vCs vx
(A.6)
where f is the friction factor (calculated from Eq. A.5). This model accounts for the effect of solid loading upon pressure loss via the total friction factor f . The pressure drop in the vortex nder can be estimated using a semi- empirical approach as, Px = 1 2 v 2 x vCS vx
2
+K
vCS vx
4 3
(A.7)
where K is the vortex nder entrance factor (K = 3.41 for rounded edge and K = 4.4 for sharp edge) The total pressure drop P = Pbody + Px can be made dimensionless Q using the average inlet velocity vin = ab leading to the so called Euler number Eu Eu = P 2 0.5vin (A.8)
254
(A.9)
where the subscript p is for particle properties and x is the particle diame xv ter. Since, for a creep ow, Rep 1 where Rep = gg rp , the drag coefcient 24 . Hence, the Stokesian drag force can be estimated as: for sphere CD = Re p Fdrag = 3g vrCS x (A.10)
where vrCS is the uniform radial gas velocity in the surface of CS given by: vrCS = Q Dx HCS (A.11)
The following expression obtained from trigonometry relations can be used to calculate HCS (R Rx )(H h) + (h S ) c R (B 2 ) (H S )
HCS
= =
if Bc > Dx if Bc Dx (A.12)
Large particles are therefore centrifuged out to the cyclone wall (because the centrifugal force is larger than the drag force. Whereas, small particles are dragged in and escape out through the vortex nder. The particle size for which the two forces balance the particles that orbit in equilibrium in CS is taken as the cyclones cut-off diameter x50 ; it is the particle size that stands a 50 - 50 chance of being captured. This particle size is called of fundamental importance and is a measure of the intrinsic separation capability of the cyclone. Here, all the gas velocity components are assumed constant over CS for the computation of the equilibrium-orbit size. Equating the forces, in Eqs. A.9 and A.10 gives the cut-off diameter x50 as: x50 = 9 vrCS g Dx 2 p vCS (A.13)
255
[pbody + px ]
(A.14)
The wall loss, or the loss in the cyclone body is given by, pbody = f AR 1.5 (vw vCS ) (0.9 Q) 2 (A.15)
where is the gas density, Q is the gas volume ow rate, AR is the total inside area of the cyclone contributing to frictional drag. It encompasses the inside area of the roof, the barrel cylinder, the cone, and the external surface of the vortex nder.
(Ht h)2 + (R Rb )2
+2Rx S (A.16)
where R = D/2, Rx = Dx /2, Rb = Bc /2. As in the Barths model [9, 77], the tangential gas velocity in the entire space between the wall and the vortex nder can be signicantly higher than the inlet velocity due to constriction of the inlet jet (see Fig. A.2). For tangential inlet, the inlet jet is compressed against the wall, resulting in a decrease in the area available for the incoming ow, and an increase in the velocity. Barth accounts for this by introducing (see Eq. A.1). 256
1+4
(1
2 )(2 1 + c0
(A.17)
where = b/R, c0 is the ratio of the mass of the incoming solids to the mass of the incoming gas in the stream feeding the cyclone. Knowing , along with vin , Rin and R, one can compute the wall velocity, vw (velocity in the vicinity of the wall), vw = (vin Rin )/(R). Now compute the geometric mean radius, Rm = Rx R which is needed in the computation of a wall axial velocity vzw . vzw = 0 .9 Q 2 ) (R2 Rm (A.18)
Trefz and Muschelknautz [174] found that, approximately 10% of the incoming gas short-circuits the cyclone and ows radially inwards in a spiral like manner along the roof and down the outside of the vortex nder. As a consequence, approximately 90% of the incoming ow Q directly participates in the ow along the walls and in the formulation of the inner vortex [77]. To calculate the friction factor f , Muschelknautz and Trefz [77] dened the cyclone body Reynolds number ReR as: ReR = Rin Rm vzw Ht (A.19)
with and are the gas phase density and absolute viscosity, respectively. Ht is the cyclone total height. The friction factor of the clean gas fair = ks ks , ReR where is the wall relative roughness of the cyclone wall f R R ks (ks = 0.046 [mm] for commercial steel pipe, < 6E 4 for non-negative R logarithm in Eq. A.22). The gas friction factor can be expressed as the sum of two components, one for smooth wall, fsm , plus an added contribution due to wall roughness fr . fair = fsm + fr
0.623 fsm = 0.323Re R
(A.20) (A.21)
257
(A.22) The total frictional drag f within a cyclone consists of two components in the MM, that due to drag on the (pure) gas phase fair and that due to an additional drag imposed by the moving strand of solids, which is present at the walls. The total friction factor f becomes, f = fair + 0.25 R Rx
0.625
fr = log
1 .6 ks 0.0005999 R
2.38 2
1 +
0.213
c0 Frx str
(A.23)
The second term in this equation is the frictional contribution due to the solids, where is the overall efciency, that is the fraction of incoming solids collected by the cyclone (Hoffmann and Stein [77] suggested to asx ) sume a value of 0.9 to 0.99). Frx is the Froude number (Frx = 2vR x g where vx is the average axial velocity through the vortex nder vx (vx = Q ), str term represents the bulk density of the dust or strand layer at R2 x the walls and can be taken as 0.4bulk where bulk is the bulk density of the solid [77]. The tangential velocity of the gas at the inner core radius Rcs (see Fig. A.3) is given by,
R Rx
vcs = vw 1+
(A.24)
R Rx
f AR 2Q
The second contribution to pressure drop is the loss in the core and in the vortex nder and is given by, px = 2 + vcs vx
2
+3
vcs vx
4/3
1 2 v 2 x
(A.25)
Cut-off diameter
A very fundamental characteristic of any lightly loaded cyclone is its cut-off diameter x50 produced by the spin of the inner vortex. This is the practical diameter that has a 50% probability of capture. The cut-off diameter is 258
Vx
0.1Q
Vcs Vz Vw
Rcs
Rc
analogous to the screen openings of an ordinary sieve or screen [77]. In lightly loading cyclones, x50 exercises a controlling inuencing on the cyclones separation performance. It is the parameter that determines the horizontal position of the cyclone grade-efciency curve (fraction collected versus particle size). For low mass loading, the cut-off diameter can be estimated in MM using Eq. A.26 [77].
x50 =
9 (0.9 Q) 2 (H S ) (p )vcs t
(A.26)
Eu = 1 + 2q 2
2 (D b) 1 +2 Dx 259
4ab 2 Dx
(A.27)
q=
2AR G ab
Dx 2(Db)
4AR G ab
0.5
(A.28)
where AR is the total wall area of the cyclone body, including the inner walls of the lid, the cylindrical and the conical sections and the outer wall of the vortex nder, given by Eq. A.29. G(= f /2 where f is the friction factor) is a wall friction factor, which Stairmand set equal to 0.005.
0.5
2 D 2 Dx (D + Bc ) AR = + Dh + Dx S + 4 2
(H h) +
D Bc 2
(A.29)
Eu =
(A.30)
(A.31)
The Ramachandran et al. [139] model was developed through a statistical analysis of pressure drop data for ninety-eight cyclone designs. The model is shown to perform better than the pressure drop models of Shepherd and Lapple [157], and Barth [9] in comparison with experimental results.
S D H h Bc D D D 1/3
Eu = 20
ab 2 Dx 260
(A.32)
Figure A.4: The control surface used in Iozia and Leith model [85].
(A.33)
where HCS is the core height (height of the control surface of Barths model) Vmax is the maximum tangential velocity, that occurs at the edge of the control surface CS , Fig. A.4. In this model however the value of the core diameter dc and the tangential velocity at the core edge Vmax are calculated from regression of experimental data using the following equations. Vmax = 6.1vin ab/D2
0.61
(Dx /D)
0.74
(Ht /D)
0.33
(A.34) (A.35)
dc = 0.52D ab/D2
0.25
(Dx /D)1.53
It is clear from this model that the most important geometry parameters that affect the cyclone collection efciency are the vortex nder diameter, the ratio of inlet area to exit area and the cyclone height. 261
This model relates, the separation cut-off diameter x50 to the pressure drop. Hence, the pressure drop needs to be predicted to use the model. A good pressure drop model for this purpose is that of Shepherd and Lapple. The interested reader can refer to Hoffmann and Stein [77] for more details.
262
B.1 Nelder-Mead
The Nelder-Mead method, also known as downhill simplex method is a commonly used nonlinear optimization technique, The technique was proposed by Nelder and Mead [119] and is a technique for minimizing an objective function in a many-dimensional space. It requires only function evaluations, and no calculation of derivatives [134]. The downhill simplex nonlinear optimization technique has been used by many researchers [e.g., 4, 12]. According to Bernon et al. [12] Powells algorithm and the downhill simplex one are ones of the most used minimization algorithms; the downhill-simplex algorithm became the most performant. Further more, the Statgraphics XV package has been used for design of experiment and optimization, with the only available optimization technique is the NelderMead technique. In this study, the target is to obtain the global optimum values. Consequently, no linear constrains applied. The idea is to employ a moving simplex in the design space to surround the optimal point and then shrink the simplex until its dimensions reach a specied error tolerance [96]. In n-dimensional space, a simplex is a gure of n +1 vertices connected by straight lines and bounded by polygonal faces. If n = 2, a simplex is a triangle; if n = 3, it is a tetrahedron. For two variables, the simplex is a triangle and the method is a pattern search that compares function values at the three vertices of a triangle. The worst vertex where f (x, y ) is largest, is rejected and replaced with a 263
new vertex. A new triangle is formed, and the search is continued. The process generates a sequence of triangles (which might have different shapes), for which the function values at the vertices get smaller and smaller. The size of the triangles is reduced, and the coordinates of the minimum point are found [108]. The algorithm is stated using the term simplex (a generalized triangle in N dimensions) and will nd the minimum of a function of N variables. It is effective and computationally compact. In the following paragraphs, a brief explanation is given for this technique. The interested reader can refer to Mathews and Fink [108].
!
R d M d W G Reection B W G Expansion d d M B R E d
Figure B.1: Basic operations in the downhill simplex method for two dimensions space [108]
to help remember that B is the best vertex, G is good (next to best), and W is the worst vertex. 264
B.1. Nelder-Mead
Shrink toward B
If the function value at C is not less than the value at W, the point G and W must be shrunk towards B (see Fig. B.1). The point G is replaced with M, and W is replaced with S, which is the midpoint of the line segment joining B with W [108].
B.2
The genetic algorithm is an optimization technique for solving both constrained and unconstrained optimization problems that is based on natural selection, the process that drives biological evolution [109]. The genetic algorithm repeatedly modies a population of individual solutions. At each step, the genetic algorithm selects individuals at random from the current population to be parents and uses them to produce the children for the next generation. Over successive generations, the population evolves toward an optimal solution. The genetic algorithm can be used to solve a variety of optimization problems that are not well suited for standard optimization algorithms, including problems in which the objective function is discontinuous, no differentiable, stochastic, or highly nonlinear [109]. The genetic algorithm uses three main types of rules at each step to create the next generation from the current population [109]: Selection rules select the individuals, called parents, that contribute to the population at the next generation. Crossover rules combine two parents to form children for the next generation. Mutation rules apply random changes to individual parents to form children. The genetic algorithm differs from a classical, derivative-based, optimization algorithm in two main ways, as summarized in Table B.1. 266
es
No
Yes Yes
If f (B) < f (R)
No
No
Replace W with R
267
Yes
No
Yes
No
Replace W with E
Replace W with R
Replace W with C
Start next cycle ... Start next cycle ... Start next cycle ... Start next cycle ...
Genetic Algorithm
It generates a population of points at each iteration. The best point in the population approaches an optimal solution. It selects the next population by computation, which uses random number generators. It can nd the global optimum solution with a high probability.
Start
Create the next generation by applying crossover & mutation (new generation: ospring)
No
Yes Stop
vector of 7 elements [27]. The population size n species how many individuals there are in each generation.
Fitness scaling
The scaling function converts raw tness scores returned by the tness function to values in a range that is suitable for the selection function. Scaling function species the function that performs the scaling. One can choose from the following functions [109]: Rank, proportional, Top or Shift linear. The rank option has been used in this study. Rank scales the raw scores based on the rank of each individual, rather than its score. The rank of an individual is its position in the sorted scores. The rank of the ttest individual is 1, the next ttest is 2, and so on. Rank tness scaling removes the effect of the spread of the raw scores. 269
Selection
Different types of selections are implanted in the optimization toolbox in Matlab 2010a [109], but only the tournament selection is used in this study (it is the only available selection type for multi-objective optimization in the Matlab toolbox). The tournament selection randomly identies some competitors from the population to compete against each other. The one with the highest performance win a parent status. The tournament selection permits to control the selective pressure put on the population. The population diversity is adjusted by modication of the competitor number. Greater competitor numbers in the tournament increase the chances to focus the search over the best individuals meaning a greater selective pressure. On the other hand, with only two competitors, the possibility of becoming a parent remains open to a larger band of the population meaning a lesser selective pressure [28]. In other words, the tournament size can take values between 1 and n (population size). Larger values give more chances to the best samples to be selected and to create offsprings. It favors a rapid, although perhaps premature, convergence to a local optimum. Very small values result in a more random selection of parents [173]. The default value of tournament size of four has been used.
Reproduction
Reproduction options determine how the genetic algorithm creates children at each new generation. Elite count species the number of individuals who are guaranteed to survive to the next generation. the Matlab manual [109] suggested to use Elite count to be two (a positive integer less than or equal to population size). Crossover fraction species the fraction of the next generation that crossover produces. Mutation produces the remaining individuals in the next generation. Matlab manual [109] suggested to use a crossover fraction of 0.8 (a fraction between 0 and 1).
Crossover
Crossover combines two individuals, or parents, to form a new individual, or child (offspring), for the next generation [109]. the Matlab optimization toolbox offers six functions: Scattered, single point, two point, intermediate, heuristic and arithmetic. The intermediate function has been used in this study. Intermediate (also, called weighted crossover [28]) creates children by a random weighted average of the parents. Intermediate crossover is controlled by a single parameter ratio: child1 = parent1+ rand*Ratio*(parent2 - parent1). If the ratio is in the range [0,1], the children produced are within the hypercube dened by the parents locations 270
Mutation
Mutation functions make small random changes in the individuals in the population, which provide genetic diversity and enable the genetic algorithm to search for a broader space. To specify the function that performs the mutation in the mutation function eld, one can choose from the following functions: constraint dependent default, gaussian, uniform and adaptive feasible. The constraint dependent default chooses: Gaussian if there are no constraints or adaptive feasible otherwise. 1) Gaussian adds a random number to each vector entry of an individual. This random number is taken from a gaussian distribution centered on zero. The standard deviation of this distribution can be controlled with two parameters. The scale parameter (default value of 1.0) determines the standard deviation at the rst generation. The shrink parameter controls how the standard deviation shrinks as generations go by. If the shrink parameter is 0, the standard deviation is constant. If the Shrink parameter is 1 (the default value), the standard deviation shrinks to 0 linearly as the last generation is reached. 2) Uniform is a two-step process. First, the algorithm selects a fraction of the vector entries of an individual for mutation, where each entry has the same probability as the mutation rate of being mutated. In the second step, the algorithm replaces each selected entry by a random number selected uniformly from the range for that entry. 3) Adaptive feasible randomly generates directions that are adaptive with respect to the last successful or unsuccessful generation. A step length is chosen along each direction so that linear constraints and bounds are satised. The constraint dependent default has been used in this study.
Stopping criteria
The stopping criteria determines what causes the algorithm to terminate. Matlab optimization toolbox has the following stopping criteria, the default values are given in parentheses. Generations specic the maximum number of iterations the genetic algorithm performs (100). Time limit species the maximum time in seconds the genetic algorithm runs before stopping (). 271
Fitness limit - If the best tness value is less than or equal to the value of tness limit, the algorithm stops (). Stall generations - If the weighted average change in the tness function value over stall generations is less than Function tolerance, the algorithm stops (50). Stall time limit - If there is no improvement in the best tness value for an interval of time in seconds specied by Stall time limit, the algorithm stops (). Function tolerance - If the cumulative change in the tness function value over stall generations is less than Function tolerance, the algorithm stops (1E-6).
B.3
Multi-objective optimization
In case of cyclone separator geometry optimization for minimum Euler number and minimum cut-off diameter, the objectives are conicting with each other. There is no best solution for which all objectives are optimal simultaneously [181]. The increase of one objective will lead to the decrease of the other objective. Then, there should be a set of solutions, the so-called Pareto optimal set or Pareto front, in which one solution cannot be dominated by any other member of this set. The denition of domination is as given in Wang et al. [181] For minimal problem, a solution a X dominates a solution b X (a b) if and only if it is superior or equal in all objectives and at least superior in one objective. This can be expressed as follows [181]: a b, if i 1, 2, . . . , m : fi (a) fi (b) j 1, 2, . . . , m : fj (a) < fj (b)
where m is the number of objective functions. Each solution of Pareto optimal set is called a non-inferior solution, which is corresponding to one point on the Pareto front. A general example of the Pareto front with two objectives is illustrated in Fig. B.4. In this example, the Pareto front is composed of six points, which are indifferent to each other (denoted by lled circles in Fig. B.4). While, points with hollow circles are not belonging to the Pareto front, since they are dominated by the Pareto front points. 272
Initially, a parent population of size n is generated randomly. All individuals in this population are sorted into different front levels based on the domination of pair comparison. Each front level is assigned a tness (or a rank) which equals its non-domination level. Level 1 is the top level in which the individual is dominated by none of the other individuals; level 2 is the secondary level in which the individual is dominated by some individuals only in level 1, and so on. In the same front level, the location of the nite number of solutions is expected to be distributed uniformly. In other words, a large diversity of the individuals can prevent the results from sticking into a local optimum. Therefore, another feature, called crowding distance, is adopted to evaluate the local aggregation of individuals. The 273
Cj =
i=1
Fij +1 Fij 1
(B.5)
where Cj is the crowding distance of point j on the Pareto front, (Fij ) is the value of the tness function i at point j (i = 1, 2 for two objective functions). For boundary points, the crowding distance is set to the maximum value of the system in order to ensure that these points can survive to the next generation. In the same rank level, individuals who have larger crowding distance also have more opportunities to be selected [182].
(B.6)
Firstly, a random number ui [0, 1] is generated, whereafter, from a specied probability distribution function, the ordinate qi is found so that the area under the probability curve from 0 to qi is equal to the chosen random number ui . The probability distribution used to create a child individual is derived to have a similar search power as that in a single-point crossover in binary-coded GAs and is given as follows [182]: 0.5( + 1) , if i 1 i (B.7) P (i ) = 0.5( + 1)/ +2 , otherwise
i
where, is the distribution index which can be any nonnegative real number. A lager value of gives a higher probability for creating near parent individuals and a small value of allows distant individuals to be selected as children individuals. Using Eq.B.7, the qi can be calculated as follows [182]: 1 (2ui ) +1 , if ui 0.5 1 (B.8) qi = +1 1 , otherwise 2(1ui ) 274
= =
(B.9)
Note that two children individuals are symmetric about the parent individuals. This is deliberately used to avoid any bias towards any particular parent individual in a single crossover operation.
Mutation Operator
Let xk be the component of an individual {xi }, which is going to be mul tated. xu k and xk stand for the maximum and minimum value of this component in all individuals, respectively. The mutated individual yk can be calculated as follows [182]:
l yk = xk + q xu k xk
(B.10)
otherwise (B.11) where m is the mutation index which is set to 20 (1- crossover fraction) in general, and r is a random number. The intermediate variables 1 and 2 stand for: 1 2 = = xk xl k l xu x k k u xk xk l xu k xk (B.12) (B.13)
Here q is a mutation parameter, which stands for: 2r + (1 2r)(1 )m +1 1/(m +1) 1, 1 q = 1 2r(1 r) + 2(r 0.5)(1 2 )m +1 1/(m +1)
if r 0.5
275
276
List of Publications
Journal Articles
Published
1. K. Elsayed, C. Lacor. Modeling and Pareto Optimization of Gas Cyclone Separator Performance Using RBF Type Articial Neural Networks and Genetic Algorithms. Powder Technology. In Press, Accepted Manuscript, Available online 17 October 2011 doi: http://dx.doi.org/10.1016/j.powtec.2011.10.015 2. K. Elsayed, C. Lacor. Numerical and Empirical Modeling of the Flow Field and Performance in Cyclones of Different Cone-Tip Diameters. Computers & Fluids, Vol. 51, No. 1, pp. 48 - 59, 2011. doi: http://dx.doi.org/10.1016/j.compfluid.2011.07.010 3. K. Elsayed, C. Lacor. Modeling, Analysis and Optimization of Aircyclones Using Articial Neural Network, Response Surface Methodology and CFD Simulation Approaches. Powder Technology, Vol. 212, No. 1, pp. 115 - 133, 2011. doi: http://dx.doi.org/10.1016/j.powtec.2011.05.002 4. K. Elsayed, C. Lacor. The Effect of Cyclone Inlet Dimensions on the Flow Pattern and Performance. Applied Mathematical Modelling, Vol. 35, No. 4, pp. 1952 - 1968, 2011. doi: http://dx.doi.org/10.1016/j.apm.2010.11.007 5. K. Elsayed, C. Lacor. Optimization of the Cyclone Separator Geometry for Minimum Pressure Drop Using Mathematical Models and CFD Simulations.Chemical Engineering Science, Vol. 65, No. 22, pp. 6048 - 6058, 2010. doi: http://dx.doi.org/10.1016/j.ces.2010.08.042 (listed in the Top 25 Hottest Articles Chemical Engineering > Chemical Engineering Science, October to December 2010).
With reviewers
1. K. Elsayed, C. Lacor. CFD Modeling and Multi-Objective Optimization of Cyclone Geometry Using Desirability Function, Articial Neural Networks and Genetic Algorithms. Applied Mathematical Modelling. Submitted, 2011. 2. K. Elsayed, C. Lacor. A CFD Study of the Effect of the Dust Outlet Geometry on the Performance and Hydrodynamics of Gas Cyclones. Computers & Fluids. Submitted, 2011. 3. K. Elsayed, C. Lacor. Numerical Study of the Effect of Cyclone Cone & Barrel Height On The Flow Pattern And Performance. Applied Mathematical Modelling. Submitted, 2011.
7. K. Elsayed, C. Lacor. Multi-Objective Optimization of Gas Cyclone Based On CFD Simulation, ECCOMAS thematic conference, CFD & Optimization, Antalya, Turkey, 23-25 May 2011. 8. K. Elsayed, C. Lacor. The effect of cyclone height on the ow pattern and performance using LES, Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3003, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 9. K. Elsayed, C. Lacor. The effect of cyclone inlet width on the ow pattern and performance, Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3085, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 10. K. Elsayed, C. Lacor. Numerical study on the effect of cyclone inlet height on the ow pattern and performance, Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG3068, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 11. K. Elsayed, C. Lacor. The effect of cyclone dustbin on the ow pattern and performance, Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3092, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 12. K. Elsayed, C. Lacor. Application of Response Surface Methodology for Modeling and Optimization of the Cyclone Separator for Minimum Pressure Drop, Fifth European Conference on Computational Fluid Dynamics (ECCOMAS CFD10), Lisbon, Portugal, 14-17 June 2010. 13. K. Elsayed, C. Lacor. The Effect of Vortex Finder Diameter on Cyclone Separator Performance and Flow Field, Fifth European Conference on Computational Fluid Dynamics (ECCOMAS CFD10), Lisbon, Portugal, 14-17 June 2010. 14. V. Agnihotri, K. Elsayed, C. Lacor, S. Verbanc. Numerical Study of Particle Deposition in the Human Upper Airways With Emphasis on Hot Spot Formation and Comparison Of LES and RANS Models, Fifth European Conference on Computational Fluid Dynamics (ECCOMAS CFD10), Lisbon, Portugal, 14-17 June 2010. 15. K. Elsayed, C. Lacor. Modeling of the Gas and Particle Flow in the Cyclone Separator Using LES, RANS and Mathematical Models, 14th International Conference on Applied Mechanics and Mechanical Engineering (AMME-14), Military Technical College, Cairo, Egypt, 2527 May 2010. 16. K. Elsayed, C. Lacor. Optimization of the Cyclone Separator Geometry Based On CFD Simulation, ERCOFTAC day, Vrije Universiteit Brussel, Brussels, Belgium, 3rd December 2009. 279
17. K. Elsayed, C. Lacor. A CFD Study of the Effects of Cone Dimensions on the Flow Field of Cyclone Separators Using LES, 13th International Conference on Aerospace Sciences & Aviation Technology (ASAT-13), Military Technical College, Cairo, Egypt, 26-28 May 2009. 18. K. Elsayed, C. Lacor. Investigation of the Geometrical Parameters Effects on the Performance and the Flow-Field of Cyclone Separators Using Mathematical Models and Large Eddy Simulation, 13th International Conference on Aerospace Sciences & Aviation Technology (ASAT-13), Military Technical College, Cairo, Egypt, 26-28 May 2009.
Technical reports
1. K. Elsayed, S. Jayaraju, C. Lacor. Wood Pellet Transport in the Biomass Burner of the Rodenhuize Power Plant, Research report, Laborelec Company, Belgium, April 2009. 2. K. Elsayed, C. Lacor. The State of the Art for Flow in Cyclone Separator, Internal report, Vrije Universiteit Brussel, Brussels, Belgium, April 2008.
280
Bibliography
[1] R. M. Alexander. Fundamentals of cyclone design and operation. In Proceedings of the Australian Institute of Mineral and Metallurgy, number 152, pages 203228, 1949. [2] M. S. M. Ali, C. J. Doolan, and V. Wheatley. Grid convergence study for a two-dimensional simulation of ow around a square cylinder at a low reynolds number. In Seventh International Conference on CFD in the Minerals and Process Industries, CSIRO, Melbourne, Australia, 9-11 December 2009. [3] S. Altmeyer, V. Mathieu, S. Jullemier, P. Contal, N. Midoux, S. Rode, and J. P. Leclerc. Comparison of different models of cyclone prediction performance for various operating conditions using a general software. Chemical Engineering and Processing, 43:511522, 2004. [4] N. Amoura, H. Kebir, S. Rechak, and J.M. Roelandt. Axisymmetric and two-dimensional crack identication using boundary elements and coupled quasi-random downhill simplex algorithms. Engineering Analysis with Boundary Elements, 34(6):611618, 2010. [5] A. Avci and I. Karagoz. Theoretical investigation of pressure losses in cyclone separators. International Communications in Heat and Mass Transfer, 28(1):107117, 1 2001. [6] Mehdi Azadi, Mohsen Azadi, and Ali Mohebbi. A CFD study of the effect of cyclone size on its performance parameters. Journal of Hazardous Materials, 182(1-3):835 841, 2010. [7] M. I. Bakari and F. Hamdullahpur. Optimization of the overall performance of cyclone separators. In 2nd Trabzon International Energy and Environment Symposium, 26-29 July, Turkey, 1998. [8] M. Balcilar, A.S. Dalkilic, and S. Wongwises. Articial neural network techniques for the determination of condensation heat transfer characteristics during downward annular ow of R134a inside a vertical smooth tube. International Communications in Heat and Mass Transfer, 38(1):75 84, 2011. 281
[9] W. Barth. Design and layout of the cyclone separator on the basis of new investigations. Brennstow-Waerme-Kraft (BWK), 8(4):19, 1956. [10] A. C. Benim, K. Ozkan, M. Cagan, and D. Gunes. Computational investigation of turbulent jet impinging onto rotating disk. International Journal of Numerical Methods for Heat & Fluid Flow, 17(3):284 301, 2007. [11] S. Bernardo, M. Mori, A.P. Peres, and R.P. Dionisio. 3-D computational uid dynamics for gas and gas-particle ows in a cyclone with different inlet section angles. Powder Technology, 162(3):190 200, 2006. [12] J. L. Bernon, V. Boudousq, J. F. Rohmer, M. Fourcade, M. Zanca, M. Rossi, and D. Mariano-Goulart. A comparative study of powells and downhill simplex algorithms for a fast multimodal surface matching in brain imaging. Computerized Medical Imaging and Graphics, 25(4):287297, 2001. [13] G. E. P. Box and K. B. Wilson. On the experimental attainment of optimum conditions. Journal of the Royal Statistical Society, 13:1 45, 1951. [14] F. Boysan, W.H. Ayer, and J. A. Swithenbank. Fundamental mathematical-modelling approach to cyclone design. Transaction of Institute Chemical Engineers, 60:222230, 1982. [15] F. Boysan, B. C. R. Ewan, J. Swithenbank, and W. H. Ayers. Experimental and theoretical studies of cyclone separator aerodynamics. IChemE Symp Series, 69:305 320, 1983. [16] D. S. Broomhead and D. Lowe. Multivariable functional interpolation and adaptive networks. Complex Systems, 2:321355, 1988. [17] H. S. Bryant, R.W. Silverman, and F. A. Zenz. How dust in gas affects cyclone pressure drop. Hydrocarbon Processing, 62(4):8790, 1983. [18] John A. Bullinaria. Introduction to neural networks, 2004. http://www.cs.bham.ac.uk/jxb/NN/l10.pdf. [19] H. Buttner. Size separation of particles from aerosol samples using impactors and cyclones. Particle & Particle Systems Characterization, 5(2):8793, 2004. [20] W. M. Carlyle, D. C. Montgomery, and G. C. Runger. Optimization problems and methods in quality control and improvement. Journal of Quality Technology, 32(1):117, 2000. [21] J. Casal and J. M. Martinez-Benet. A better way to calculate cyclone pressure drop. Chemical Engineering, 90(2):99100, 1983. 282
[22] J. Chen and M. Shi. A universal model to calculate cyclone pressure drop. Powder Technology, 171(3):184 191, 2007. [23] T. G. Chuah, J. Gimbun, and T. S.Y. Choong. A CFD study of the effect of cone dimensions on sampling aerocyclones performance and hydrodynamics. Powder Technology, 162:126 132, 2006. [24] T. J. Chung. Computational Fluid Dynamics. Cambridge University Press, 1st edition, 2002. [25] C. A. C. Coello. Twenty years of evolutionary multi-objective optimization: a historical view of the eld. IEEE Computational Intelligence Magazine, 1:2836, 2006. [26] C. A. C. Coello, G. B. Lamont, and D. A. Van Veldhuizen. Evolutionary Algorithms for Solving Multi-Objective Problems. Springer, 2nd edition, 2007. [27] D. Cooley. Using optimtool GA in Matlab, 2011. digital.cs.usu.edu/cooley/CS6665_10_Optimtool_GA.pptx. [28] G. Corriveau, R. Guilbault, and A. Tahan. Genetic algorithms and nite element coupling for mechanical optimization. Advances in Engineering Software, 41(3):422 426, 2010. [29] C. Cort es and A. Gil. Modeling the gas and particle ow inside cyclone separators. Progress in Energy and Combustion Science, 33(5):409 452, 2007. [30] J. S. Cowpe, J. S. Astin, R. D. Pilkington, and A. E. Hill. Application of response surface methodology to laser-induced breakdown spectroscopy: inuences of hardware conguration. Spectrochimica Acta Part B, 62:1335 1342, 2007. [31] J. S. Curtis and B. V. Wachem. Modeling particle-laden ows: A research outlook. Wiley InterScience, 50-11, 2004. [32] L. Davidson. Lecture notes. Dept. of Thermo and Fluid Dynamics, Chalmers University of Technology, G oteborg, Sweden, 2000. [33] J. W. Deardorff. The use of subgrid transport equations in a threedimensional model of atmospheric turbulence. Journal of Fluids Engineering, 95:429438, 1973. [34] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transaction Evolitionery Computation, 6:182197, 2002. [35] W. Deconinck. Design and application of discrete explicit lters for large eddy simulation of compressible turbulent ows. Masters thesis, University of Toronto, 2008. 283
[36] J. J. Derksen. Separation performance predictions of a stairmand high-efciency cyclone. AIChE Journal, 49(6):13591371, 2003. [37] J. J. Derksen, S. Sundaresan, and H. E. A. van den Akker. Simulation of mass-loading effects in gassolid cyclone separators. Powder Technology, 163:5968, 2006. [38] J. J. Derksen and H. E. A. Van den Akker. Simulation of vortex core precession in a reverse-ow cyclone. AIChE Journal, 46(7):1317 1331, 2000. [39] J. J. Derksen, H. E. A. van den Akker, and S. Sundaresan. Twoway coupled large-eddy simulations of the gas-solid ow in cyclone separators. AIChE Journal, 54(4):872885, April 2008. [40] D. B. Dias, M. Mori, and W. P. Martignoni. Boundary condition effects in CFD cyclone simulations. In 8th World Congress of Chemical Engineering (WCCE8), Montreal, August 2009. [41] K. K. J. R. Dinesh and M. P. Kirkpatrick. Study of jet precession, recirculation and vortex breakdown in turbulent swirling jets using LES. Computers & Fluids, 38:12321242, 2009. [42] J. Dirgo. Relationship between cyclone dimensions and performance. PhD thesis, Harvard University, USA, 1988. [43] J. Dirgo and D. Leith. Cyclone collection efciency: Comparison of experimental results with theoretical predictions. Aerosol Science and Technology, 4:401415, 1985. [44] S. E. Elghobashi. On predicting particle-laden turbulent ows. Applied Scientic Research, 52:309329, 1994. [45] K. Elsayed and C. Lacor. Modeling and pareto optimization of gas cyclone separator performance using RBF type articial neural networks and genetic algorithms. Powder Technology, Accepted 9 October, 2011. http://dx.doi.org/10.1016/j.powtec.2011.10.015. [46] K. Elsayed and C. Lacor. Investigation of the geometrical parameters effects on the performance and the ow-eld of cyclone separators using mathematical models and large eddy simulation. In 13th Aerospace Sciences & Aviation Technology (ASAT-13), Military Technical College, Cairo, Egypt, 26-28 May 2009. [47] K. Elsayed and C. Lacor. The effect of cyclone dustbin on the ow pattern and performance. In Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3092, 2010, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 284
[48] K. Elsayed and C. Lacor. The effect of cyclone height on the ow pattern and performance using LES. In Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3003, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. [49] K. Elsayed and C. Lacor. The effect of vortex nder diameter on cyclone separator performance and ow eld. In Fifth European Conference on Computational Fluid Dynamics (ECCOMAS CFD 2010), Lisbon, Portugal, 14-17 June 2010. [50] K. Elsayed and C. Lacor. Optimization of the cyclone separator geometry for minimum pressure drop using mathematical models and CFD simulations. Chemical Engineering Science, 65(22):60486058, 2010. [51] K. Elsayed and C. Lacor. A CFD study of the effect of cyclone barrel height on its performance parameters. In 8th International Conference on Computational Fluid Dynamics in the Oil & Gas, Metallurgical and Process Industries, Trondheim, Norway, 21-23 June 2011. [52] K. Elsayed and C. Lacor. The effect of cyclone inlet dimensions on the ow pattern and performance. Applied Mathematical Modelling, 35(4):19521968, 2011. [53] K. Elsayed and C. Lacor. Modeling, analysis and optimization of aircyclones using articial neural network, response surface methodology and CFD simulation approaches. Powder Technology, 212(1):115133, 2011. [54] K. Elsayed and C. Lacor. Multi-objective optimization of gas cyclone based on CFD simulation. In ECCOMAS thematic conference, CFD & Optimization, Antalya, Turkey, 23-25 May 2011. [55] K. Elsayed and C. Lacor. Numerical modeling of the ow eld and performance in cyclones of different cone-tip diameters. Computers & Fluids, 51(1):4859, 2011. [56] K. Elsayed and C. Lacor. Single and multi-objective optimization of the cyclone separator geometry using articial neural network and genetic algorithm. In EUROGEN 2011, Evolutionary and Deterministic Methods for Design, Optimization and Control with Applications to Industrial and Societal Problems, Capua, Italy, 14-16 September, 2011. [57] Ferit Ficici, Vedat Ari, and Murat Kapsiz. The effects of vortex nder on the pressure drop in cyclone separators. International Journal of the Physical Sciences, 5(6):804813, 2010. [58] M. W. First. Cyclone dust collector design. ASME Annual General Meeting, Paper No. 49A127 (1949). 285
[59] Fluent. FLUENT 6.3 users guide. Fluent Incorporated, Lebanon, 2006. [60] M. Germano, U. Piomelli, P. Moin, and W. H. Cabot. A dynamic subgrid-scale eddy viscosity model. Physics of Fluids A, 3(7):1760 1765, 1991. [61] M. Gfrerer and E. Lankmayr. Screening, optimization and validation of microwave-assisted extraction for the determination of persistent organochlorine pesticides. Analytica Chimica Acta, 533:203 211, 2005. [62] J. Gimbun, T. G. Chuah, T. S. Y. Choong, and A. Fakhrul-Razi. A CFD study on the prediction of cyclone collection efciency. International Journal for Computational Methods in Engineering Science and Mechanics, 6(3):161 168, 2005. [63] J. Gimbun, T.G. Chuah, T.S. Choong, and Y.A. Fakhrul-Razi. Prediction of the effects of cone tip diameter on the cyclone performance. Aerosol Science and Technology, 36:10561065, 2005. [64] J. Gimbun, T.G. Chuah, A. Fakhrul-Razi, and T. S. Y. Choong. The inuence of temperature and inlet velocity on cyclone pressure drop: a CFD study. Chemical Engineering & Processing, 44(1):712, 2005. [65] F. Girosi, T. Poggio, and A. Caprile. Extensions of a theory of networks for approximations and learning: outliers and negative examples. In NIPS-3: Proceedings of the 1990 conference on Advances in neural information processing systems 3, pages 750756, Denver, Colorado, United States, 1990. [66] A. Gorton-Hulgerth, J. Woisetschlager, Graham Wigley, and G. Staudinger. Investigation of the ow eld in the upper part of a cyclone with laser and phase doppler anemometry. Particle & Particle Systems Characterization, 17:2127, 2000. [67] A. D. Gosman and E. Ioannides. Aspects of computer simulation of liquid-fuelled combustor. AIAA Journal, 81-0323, 1981. [68] W. D. Grifths and F. Boysan. Computational uid dynamics (CFD) and empirical modelling of the performance of a number of cyclone samplers. Journal of Aerosol Science, 27(2):281304, 1996. [69] G. Gronald and J. J. Derksen. Simulating turbulent swirling ow in a gas cyclone: A comparison of various modeling approaches. Powder Technology, 205(1-3):160 171, 2011. [70] P. O. Haaland. Experimental design in biotechnology. Marcel Dekker, New York, 1989. 286
[71] E. C. Jr. Harrington. The desirability function. Industrial Quality Control, 21:494498, 1965. [72] Eric Hartman, James D. Keeler, and Jacek M. Kowalski. Layered neural networks with gaussian hidden units as universal approximations. Neural Computation, 2(2):210215, 1990. [73] J. O. Hinze. Turbulence. McGraw-Hill, New York, 2nd ed. edition, 1975. [74] A. J. Hoekstra. Gas ow eld and collection efciency of cyclone separators. PhD thesis, Technical University Delft, 2000. [75] A. J. Hoekstra, J. J. Derksen, and H. E. A. Van Den Akker. An experimental and numerical study of turbulent swirling ow in gas cyclones. Chemical Engineering Science, 54:20552065, 1999. [76] A. C. Hoffmann, M. de Groot, W. Peng, H. W. A. Dries, and J. Kater. Advantages and risks in increasing cyclone separator length. AIChE Journal, 47(11):24522460, 2001. [77] A. C. Hoffmann and L. E. Stein. Gas cyclones and swirl tubes: Principle, Design and Operation. Springer, 2nd edition, 2008. [78] A.C. Hoffmann, M. De Groot, and A. Hospers. The effect of the dust collection system on the ow pattern and separation efciency of a gas cyclone. The Canadian Journal of Chemical Engineering, 74:464470, August 1996. [79] A. Horvath, C. Jordan, and M. Harasek. Inuence of vortex-nder diameter on axial gas ow in simple cyclone. Chemical Product and Process Modeling, 3(1), 2008. [80] Y. H. Hu and J. Hwang, editors. Handbook of Neural Network Signal Processing. CRC Press, Washington, D.C., 2002. [81] S. Huang, Q. S. Li, and S. Xu. Numerical evaluation of wind effects on a tall steel building by CFD. Journal of Constructional Steel Research, 63(5):612 627, 2007. [82] S. H. Huang, Q. S. Li, and J. R. Wu. A general inow turbulence generator for large eddy simulation. Journal of Wind Engineering and Industrial Aerodynamics, 98(10-11):600 617, 2010. [83] P. Hutchinson, G. F. Hewitt, and A. E. Dukler. Deposition of liquid or solid dispersions from turbulent gas streams: a stochastic model. Chemical Engineering Science, 26:419439, 1971. [84] D. L. Iozia and D. Leith. Effect of cyclone dimensions on gas ow pattern and collection efciency. Aerosol Science and Technology, 10(3):491500, 1989. 287
[85] D. L. Iozia and D. Leith. The logistic function and cyclone fractional efciency. Aerosol Science and Technology, 12(3):598606, 1990. [86] S. T. Jayaraju. Study of the air ow and aerosol transport in the human upper airway using LES and DES methodology. PhD thesis, Vrije Universiteit Brussel, 2009. [87] Zhang Jian and Jin You-hai. CFD analysis on collection performance of swirl tube with different vortex nder geometries. Journal of System Simulation, 21(4):121417, 2009. [88] Viviana P. Jofre, Mariela V. Assof, Martn L. Fanzone, Hctor C. Goicoechea, Luis D. Martnez, and Mara F. Silva. Optimization of ultrasound assisted-emulsication-dispersive liquid-liquid microextraction by experimental design methodologies for the determination of sulfur compounds in wines by gas chromatography-mass spectrometry. Analytica Chimica Acta, 683(1):126 135, 2010. [89] W. P. Jones and M. Wille. Large-eddy simulation of a plane jet in a cross-ow. International Journal of Heat and Fluid Flow, 17-3:296 306, 1995. [90] I. Karagoz and A. Avci. Modelling of the pressure drop in tangential inlet cyclone separators. Aerosol Science and Technology, 39(9):857 865, 2005. [91] F. Kaya and I. Karagoz. Performance analysis of numerical schemes in highly swirling turbulent ows in cyclones. Current Science, 94(10):1273 1278, 2008. [92] F. Kaya and I. Karagoz. Numerical investigation of performance characteristics of a cyclone prolonged with a dipleg. Chemical Engineering Journal, 151:3945, 2009. [93] L. C. Kenny and R. A. Gussman. A direct approach to the design of cyclones for aerosol-monitoring applications. Journal of Aerosol Science, 31(12):14071420, 2000. [94] Abolfazl Khalkhalia and Hamed Sakhania. Pareto based objective optimization of a cyclone vortex nder using GMDH type neural networks and genetic algorithms. neering Optimization, , Available online: 05 Jul 2011, http://dx.doi.org/10.1080/0305215X.2011.564619. multiCFD, Engi2011.
[95] J. C. Kim and K. W. Lee. Experimental study of particle collection by small cyclones. Aerosol Science and Technology, 12:10031015, 1990. [96] J. Kiusalaas. Numerical Methods in Engineering With MATLAB. Cambridge University Press, Cambridge, UK, 2nd edition, 2010. 288
[97] C. Lacor. Lecture notes. Dept. Fluid Mechanics, Vrije Universiteit Brussel, Brussels, Belgium, 2007. [98] B. E. Launder, G. J. Reece, and W. Rodi. Progress in the development of a Reynolds-stress turbulence closure. Journal of Fluid Mechanics, 68(03):537566, 1975. [99] D. Leith and W. Licht. The collection efciency of cyclone type particle collectors: a new theoretical approach. A.I.Ch.E. Symposium Series, 68:196206, 1972. [100] A. Leonard. Energy cascade in LES of turbulent uid ows. Advances in Geophysics, 18:237248, 1974. [101] D. K. Lilly. A proposed modication of the germano subgrid-scale closure method. Physics of Fluids A, 4:633635, 1992. [102] K. S. Lim, H. S. Kim, and K. W. Lee. Characteristics of the collection efciency for a cyclone with different vortex nder shapes. Journal of Aerosol Science, 35(6):743754, 2004. [103] K. S. Lim, S. B. Kwon, and K. W. Lee. Characteristics of the collection efciency for a double inlet cyclone with clean air. Journal of Aerosol Science, 34(8):10851095, 2003. [104] A. J. Linden. Investigations into cyclone dust collectors. Proceedings of the Institution of Mechanical Engineers Part B Engineering Manufacture (Professional Engineering Publishing), 160:233251, 1949. [105] Z. Liu, Y. Zheng, L. Jia, J. Jiao, and Q. Zhang. Stereoscopic PIV studies on the swirling ow structure in a gas cyclone. Chemical Engineering Science, 61:4252 4261, 2006. [106] C. Liyana-Pathirana and F. Shahidi. Optimization of extraction of phenolic compounds from wheat using response surface methodology. Food Chemistry, 93:4756, 2005. [107] P. Worth Longest and Samir Vinchurkar. Effects of mesh style and grid convergence on particle deposition in bifurcating airway models with comparisons to experimental data. Medical Engineering & Physics, 29(3):350 366, 2007. [108] J. H. Mathews and K. D. Fink. Numerical Methods Using MATLAB. Prentice Hall, 3rd edition, 1999. [109] Matlab. Global Optimization Toolbox Users Guide, Matlab R2010a. The MathWorks, Inc., 2010. [110] P. Moin and J. Kim. Numerical investigation of turbulent channel ow. Journal of Fluid Mechanics, 155:441, 1982. 289
[111] M. E. Moore and A. R. McFarland. Performance modeling of singleinlet aerosol sampling cyclones. Environmental Science and Technology, 27(9):18421848, September 1993. [112] S. A. Morsi and A. J. Alexander. An investigation of particle trajectories in two-phase ow systems. Journal of Fluid Mechanics, 55(02):193208, 1972. [113] S. Murakami. Overview of turbulence models applied in CWE-1997. Journal of Wind Engineering and Industrial Aerodynamics, 74-76:1 24, 1998. [114] E. Muschelknautz. Die berechnung von zyklonabscheidern fur gas. Chemie Ingenieur Technik, 44:6371, 1972. [115] E. Muschelknautz and W. Kambrock. Aerodynamische beiwerte des zyclon abscheiders aufgrund neuer und verbesserter messungen. Chem-Ing-Tech, 42(5):24755, Chemie Ingenieur Technik 1970. [116] E. Muschelknautz and M. Trefz. Design and calculation of higher and design and calculation of higher and highest loaded gas cyclones. in Proceedings of Second World Congress on Particle Technology Kyoto, Japan, pages 5271, October 1990. [117] E. Muschelknautz and M. Trefz. VDI-Warmeatlas. Druckverlust und Abscheidegrad in Zyklonen, 6 Auage:Lj1 Lj9, 1991. [118] Raymond H. Myers, Douglas C. Montgomery, and Christine M. Anderson-Cook. Response Surface Methodology: Process and Product Optimization Using Designed Experiments. John Wiley & Sons, Inc., 3rd edition, 2009. [119] J. A. Nelder and R. Mead. A simplex method for function minimization. The Computer Journal (1965), 7(4). [120] Antonios Niros and George Tsekouras. A fuzzy clustering algorithm to estimate the parameters of radial basis functions neural networks and its application to system modeling. In John Darzentas, George Vouros, Spyros Vosinakis, and Argyris Arnellos, editors, Articial Intelligence: Theories, Models and Applications, volume 5138 of Lecture Notes in Computer Science, pages 194204. Springer Berlin / Heidelberg, 2008. [121] Stefan Obermair, Christoph Gutschi, Jakob Woisetschlager, and Gernot Staudinger. Flow pattern and agglomeration in the dust outlet of a gas cyclone investigated by phase doppler anemometry. Powder Technology, 156:3442, 2005. [122] Stefan Obermair and Gernot Staudinger. The dust outlet of a gas cyclone and its effects on separation efciency. Chemical Engineering Technology, WILEY-VCH Verlag GmbH, 24(12):12591263, 2001. 290
[123] Stefan Obermair, Jakob Woisetschlager, and Gernot Staudinger. Investigation of the ow pattern in different dust outlet geometries of a gas cyclone by laser doppler anemometry. Powder Technology, 138:239251, 2003. [124] T. ODoherty, A. J. Grifths, N. Syred, P. J. Bowen, and W. Fick. Experimental analysis of rotating instabilities in swirling and cyclonic ows. Developments in Chemical Engineering and Mineral Processing, 7:245267, 1999. [125] T. J. Overcamp and S. E. Scarlett. Effect of Reynolds number on the Stokes number of cyclones. Aerosol Science and Technology, 19(3):362370, 1993. [126] S. H. R. Pasandideh and S. T. A. Niaki. Multi-response simulation optimization using genetic algorithm within desirability function framework. Applied Mathematics and Computation, 175(1):366 382, 2006. [127] W. Peng, A. C. Hoffmann, P. J. A. J. Boot, A. Udding, H.W.A. Dries, A. Ekker, and J. Kater. Flow pattern in reverse-ow centrifugal separators. Powder Technology, 127:212 222, 2002. [128] R. H. Perry and D. W. Green, editors. Perrys Chemical Engineers Handbook. McGraw-Hill, 1997. [129] U. Piomelli, Y. Yu, and R. J. Adrain. Sub-grid scale energy transfer and near-wall turbulence structure. Physics of Fluids, 8:215 224, 1996. [130] S. I. Pishbin and M. Moghiman. Optimization of cyclone separators using genetic algorithm. International Review of Chemical Engineering (I.RE.CH.E.), 2(6):683690, 2010. [131] N. W. Polhemus. How to perform an optimization experiment using statgraphics centurion, 2005. http://www.sigmaplus.fr/telecharg/StatG/howto9.pdf. [132] S. B. Pope. Turbulent ows. Cambridge university press, New York, 2000. [133] L. Prandtl. Bericht uber die entstehung der turbulenz. Zeitschrift Angewandte Mathematik und Mechanik, 5:136139, 1925. fur [134] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. Numerical Recipes in FORTRAN 77: The Art of Scientic Computing. Cambridge University Press, Cambridge, UK, 1992. [135] William H. Press, Brian P. Flannery, Saul A. Teukolsky, and William T. Vetterling. Numerical Recipes in C The Art of Scientic Computing. Cambridge University Press, Cambridge, UK, 1992. 291
[136] F. Qian and Y. Wu. Effects of the inlet section angle on the separation performance of a cyclone. Chemical Engineering Research and Design, 87(12):15671572, 2009. [137] F. Qian and M. Zhang. Effects of the inlet section angle on the ow eld of a cyclone. Chemical Engineering & Technology, 30(11):1521 4125, 2007. [138] Fuping Qian, Jiguang Zhang, and Mingyao Zhang. Effects of the prolonged vertical tube on the separation performance of a cyclone. Journal of Hazardous Materials, B(136):822829, 2006. [139] G. Ramachandran, D. Leith, J. Dirgo, and H. Feldman. Cyclone optimization based on a new empirical model for pressure drop. Aerosol Science and Technology, 15:135148, 1991. [140] A. Raou, M. Shams, M. Farzaneh, and R. Ebrahimi. Numerical simulation and optimization of uid ow in cyclone vortex nder. Chemical Engineering and Processing, 47:128137, 2008. [141] G. Ravi, Santosh K. Gupta, and M. B. Ray. Multiobjective optimization of cyclone separators using genetic algorithm. Ind. Eng. Chem. Res., 39:42724286, 2000. [142] K. Rietema. Het mechanisme van de afscheiding van jnverdeelde stoffen in cyclonen (in dutch). De Ingenieur, 71(39):ch59ch65, 1959. [143] P. J. Roache. Perspective: A method for uniform reporting of grid renement studies. Journal of Fluids Engineering, 116(3):405413, 1994. [144] Patrick J. Roache. Verication of codes and calculations. AIAA Journal, 36(5):696702, 1998. [145] PJ Roache. Quantication of uncertainty in computational uid dynamics. Annual Review of Fluid Mechanics, 29:123160, 1997. [146] H. Sakhani, M. A. Akhavan-Behabadi, N. Nariman-Zadeh, and M. J. Mahmood Abadi. Modeling and multi-objective optimization of square cyclones using CFD and neural networks. Chemical Engineering Research and Design, 89(3):301 309, 2011. [147] H. Sakhani, M.A. Akhavan-Behabadi, M. Shams, and M. H. Rahimyan. Numerical simulation of ow eld in three types of standard cyclone separators. Advanced Powder Technology, 21(4):435 442, 2010. [148] H. Sakhani, A. Hajiloo, M.A. Ranjbar, and N. Nariman-Zadeh. Modeling and multi-objective optimization of cyclone separators using CFD and genetic algorithms. Computers & Chemical Engineering, 35(6):10641071, 2011. 292
[149] H. Sakhani, S.A. Nourbakhsh, and N. Nariman-zadeh. Modeling and multi-objective optimization of cyclone vortex nder using CFD and neural networks. In 2nd International Conference on Engineering Optimization, September 6-9, 2010, Lisbon, Portugal, pages 19, 2010. [150] P. Sagaut. Large-Eddy Simulation for Incompressible Flows. Scientic Computation, Springer, 2002. [151] B.E. Saltzman and J.M. Hochstrasser. Design and performance of miniature cyclone for respirable aerosol sampling. Environmental Science & Technology, 17(7):418424, July 1983. [152] S. Schmidt, H. M. Blackburn, M. Rudman, and I. Sutalo. Simulation of turbulent ow in a cyclonic separator. In Third international conference on CFD in the minerals and process industries, CSIRO, Melbourne, Australia, 10-12 December 2003. [153] St. Schmidt, H. M. Blackburn, and M. Rudman. Impact of outlet boundary conditions on the ow properties within cyclone. In 15th Australasian Fluid Mechanics Conference, The University of Sydney, Sydney, Australia, 2004. [154] H. Shalaby. On the Potential of Large Eddy Simulation to Simulate Cyclone Separators. PhD thesis, Chemnitz University of Technology, Chemnitz, Germany, 2007. [155] H. Shalaby, K. Pachler, K. Wozniak, and G. Wozniak. Comparative study of the continuous phase ow in a cyclone separator using different turbulence models. International Journal for Numerical Methods in Fluids, 48(11):11751197, 2005. [156] H. Shalaby, K. Wozniak, and G. Wozniak. Particle-laden ow simulation in a cyclone separator. WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim, 6(1):547548, 2006. [157] C. B. Shepherd and C. E. Lapple. Flow pattern and pressure drop in cyclone dust collectors cyclone without intel vane. Industrial & Engineering Chemistry, 32(9):12461248, September 1940. [158] Sujeet Kumar Shukla, Prashant Shukla, and Pradyumna Ghosh. Evaluation of numerical schemes using different simulation methods for the continuous phase modeling of cyclone separators. Advanced Powder Technology, 22(2):209219, 2010. [159] M. D. Slack, R. O. Prasad, A. Bakker, and F. Boysan. Advances in cyclone modeling using unstructured grids. Trans IChemE., 78 Part A, (2000), 2000. [160] John W. Slater. Examining spatial (grid) convergence, 2008. http://www.grc.nasa.gov/WWW/wind/valid/tutorial/spatconv.html . 293
[161] J. Smagorinsky. General circulation experiments with the primitive equations: The basic experiment. Monthly weather review, 91(3):99 164, 1963. [162] A. Smirnov, S. Shi, and I. Celik. Random ow generation technique for large eddy simulations and particle-dynamics modeling. Journal of Fluids Engineering, 123(2):359371, 2001. [163] G. Solero and A. Coghe. Experimental uid dynamic characterization of a cyclone chamber. Experimental Thermal and Fluid Science, 27:8796, 2002. [164] M. Sommerfeld, A. Ando, and D. Wennerberg. Swirling, particleladen ows through a pipe expansion. Journal of Fluids Engineering, 114:648656, 1992. [165] C. J. Stairmand. Pressure drops in cyclone separators. Industrial and Engineering Chemistry, 16(B):409411, 1949. [166] C. J. Stairmand. The design and performance of cyclone separators. Industrial and Engineering Chemistry, 29:356383, 1951. [167] G. Steenackers, R. Versluys, M. Runacres, and P. Guillaume. Reliability-based design optimization of computation-intensive models making use of response surface models. Quality and Reliability Engineering International, 27(4):555568, 2011. [168] Gunther Steenackers. Finite Element Model Updating and Optimization of Mechanical Systems Making Use of Regressive Techniques. PhD thesis, Vrije Universiteit Brussel, 2008. [169] Gunther Steenackers and Patrick Guillaume. Bias-specied robust design optimization: A generalized mean squared error approach. Computers & Industrial Engineering, 54(2):259 268, 2008. [170] Gunther Steenackers, Patrick Guillaume, and Steve Vanlanduit. Robust optimization of an airplane component taking into account the uncertainty of the design parameters. Quality and Reliability Engineering International, 25(3):255282, 2009. [171] A.C. Stern, K.J. Caplan, and P.D. Bush. Cyclone dust collectors. New York: American Petroleum Institute, 1955. [172] P. K. Swamee, N. Aggarwal, and K. Bhobhiya. Optimum design of cyclone separator. American Institute of Chemical Engineers (AIChE ), 55(9):22792283, 2009. [173] Dominique Thevenin and Gabor Janiga. Optimization and computational uid dynamics. Springer Verlag, Berlin, 2008. 294
[174] M. Trefz. Die vershiedenen abscheidevorgange im hoher un hoch beladenen gaszyklon unter besonderer berucksichtigung der sekundarstromung. Forschritt-Berichte VDI; VDI-Verlag GmbH: Dusseldorf, Germany (1992), 295. [175] M. Trefz and E. Muschelknautz. Extended cyclone theory for gas ows with high solids concentrations. Chemie Ingenieur Technik, 16:153160, 1993. [176] P. Venkataraman. Applied optimization with MATLAB programming. John Wiley and Sons, 2nd edition, 2009. [177] H.K. Versteeg and W. Malalasekera. Introduction to Computational Fluid Dynamics, the Finite Volume Method. Longman Scientic & Technical, 1995. [178] B. Wang, D.L. Xu, K.W. Chu, and A.B. Yu. Numerical study of gassolid ow in a cyclone separator. Applied Mathematical Modelling, 30:13261342, 2006. [179] B. Wang, D.L. Xu, G. X. Xiao, K. W. Chu, and A. B. Yu. Numerical study of gas-solid ow in a cyclone separator. In Third International Conference on CFD in the Minerals and Process Industries, pages 371376, CSIRO, Melbourne, Australia, December 2003. [180] Lei Wang, Cheng Shao, Hai Wang, and Hong Wu. Radial basis function neural networks-based modeling of the membrane separation process: Hydrogen recovery from renery gases. Journal of Natural Gas Chemistry, 15:230234, 2006. [181] X.D. Wang, C. Hirsch, Sh. Kang, and C. Lacor. Multi-objective optimization of turbomachinery using improved NSGA-II and approximation model. Computer Methods in Applied Mechanics and Engineering, 200(9-12):883 895, 2011. [182] Xiaodong Wang. CFD Simulation of Complex Flows in Turbomachinery and Robust Optimization of Blade Design. PhD thesis, Vrije Universiteit Brussel, 2010. [183] D. C. Wilcox. Turbulence Modeling for CFD. DCW Industries, Inc, 2nd ed. edition, 1994. [184] R. Xiang, S. H. Park, and K. W. Lee. Effects of cone dimension on cyclone performance. Journal of Aerosol Science, 32(4):549561, 2001. [185] R. B. Xiang and K. W. Lee. Numerical simulation of ow patterns in cyclones of different cone dimensions. Part. Part. Syst. Charact., 22(3):212218, 2005. 295
[186] R. B. Xiang and K. W. Lee. Numerical study of ow eld in cyclones of different height. Chemical Engineering and Processing, 44:877 883, 2005. [187] H. Yoshida, Y. Inada, K. Fukui, and T. Yamamoto. Improvement of gas-cyclone performance by use of local uid ow control method. Powder Technology, 193:614, 2009. [188] Jin You-hai, Ji Guang-qin, Cao Qing-yun, and Wang Jian-jun. Numerical simulation of gas-phase ow eld in vortex nder of cyclone separators. Journal of China University of Petroleum, 32(6):10912, 2008. [189] Y. Yuan, Y. Gao, L. Mao, and J. Zhao. Optimisation of conditions for the preparation of -carotene nanoemulsions using response surface methodology. Food Chemistry, 107:13001306, 2008. [190] Aishe Zhang and Ling Zhang. RBF neural networks for the prediction of building interference effects. Computers & Structures, 82(27):23332339, 2004. [191] B. Zhang and S. Hui. Numerical simulation and PIV study of the turbulent ow in a cyclonic separator. In International Conference on Power Engineering, Hangzhou, China, October 2007. [192] B. Zhao. A theoretical approach to pressure drop across cyclone separators. Chemical Engineering Technology, 27:11051108, 2004. [193] B. Zhao. Experimental investigation of ow patterns in cyclones with conventional and symmetrical inlet geometries. Chemical Engineering & Technology, 28(9):969 972, 2005. [194] B. Zhao. Modeling pressure drop coefcient for cyclone separators a support vector machine approach. Chemical Engineering Science, 64:41314136, 2009. [195] B. Zhao, H. Shen, and Y. Kang. Development of a symmetrical spiral inlet to improve cyclone separator performance. Powder Technology, 145:4750, 2004. [196] B. Zhao and Y. Su. Particle collection theory for cyclone separators: summary and comparison. Particle & Particle Systems Characterization, 23:484488, 2006. [197] B. Zhao and Y. Su. Articial neural network-based modeling of pressure drop coefcient for cyclone separators. chemical engineering research and design, 88:606613, 2010. [198] B. Zhao, Y. Su, and J. Zhang. Simulation of gas ow pattern and separation efciency in cyclone with conventional single and spiral double inlet conguration. Chemical Engineering Research and Design, 84:11581165, 2006. 296
[199] Y. Zhu, M. C. Kim, K. W. Lee, Y. O. Park, and M. R. Kuhlman. Design and performance evaluation of a novel double cyclone. Aerosol Science and Technology, 34(8):373380, 2000. [200] Y. Zhu and K. W. Lee. Experimental study on small cyclones operating at high ow rates. Journal of Aerosol Science, 30:13031315, 1999.
297