Proposed method to obtain initial guess values of model parameters to identify FOPTD transfer function models of MIMO systems using closed-loop step responses by optimization method.
Proposed method to obtain initial guess values of model parameters to identify FOPTD transfer function models of MIMO systems using closed-loop step responses by optimization method.
Proposed method to obtain initial guess values of model parameters to identify FOPTD transfer function models of MIMO systems using closed-loop step responses by optimization method.
Dr. M. Chidambaram Department of Chemical Engg IIT-Madras chidam@iitm.ac.in 1 Process Identification Most chemical processes are multi-input and multi-output (MIMO) systems.
Transfer function model matrix required to design multivariable controllers.
For controller tuning purposes, closed loop identification method is preferred over open loop method (Hjalmarsson et al., 1996). 2 3 Viswanathan et al.(2001) Proposed GA+BFGS optimization method for identification, in time domain Computational time is high (about 1 hr and 30 min). Parameter convergence not guaranteed. Upper and lower bounds taken arbitrarily. Objectives To propose a simple method to obtain initial guess values of model parameters to identify FOPTD transfer function models of MIMO systems using closed-loop step responses by optimization method.
Extend this method to identify SOPTD transfer function model identification for multivariable systems. Model Identification Procedure 4 TITO Process Transfer function matrix Decentralized controller matrix
Response matrix of TITO system 5 Objective Function for Optimization Guess values for FOPTD model parameters Design value of is taken as 6 Diagonal process gains obtained from controller gain. The value of Time delay considered as same as closed time delay.
Time constant obtained from settling time. Guess value for time constant is assumed as Approximated FOPTD Model Guess values of non-diagonal process gains are estimated from closed-loop transfer function. 7 The value of S* is taken as 8 Three simulation examples considered to show the effectiveness of the proposed method. 9 9 9 Simulation examples Identification
Identification of FOPTD to FOPTD TITO system. Effect of different controller settings, noise & initial guess values are studied. Optimization computational time considerably reduced from 1 hr and 29 mins 1 to 4 min and 17 s Identification of Higher order model to FOPTD TITO system. Effect of noise & initial guess values are studied. Optimization computational time considerably reduced from 1 hr and 20 mins 1 to 1 min and 26 s. Identification of SOPTD to SOPTD 3 3 system. Effect of controller settings & initial guess values are studied. Optimization computational time is 23 min and 20 s Example 1.1: Wood and Berry (1983)
distillation process. Simulation Examples Decentralized PID controller settings(Ham & Kim, 1998) 10 Converged model parameters match well with actual FOPTD system.
Effect of different controller settings, noise & initial guess values are studied.
Optimization computational time considerably reduced from 1 hr and 29 mins 1 to 4 min and 17 s. Closed-loop response of the process and the identified FOPTD transfer function model using decentralized PID controller.(Model & Process response curves overlap) 11 Table 1.1. Model identification results using decentralized PID controller for Example 1. Model IAE values ISE values Time (N. I.) Model identification results without noise 11 Guess 5.0505 6.2500 1.0000 6.4110 -4 1.1710 -8
Converged -19.1683 14.0428 2.9041 12 N.I: Number of iterations (in brackets); Time in Seconds 13 Example 1.2: Niederlinski (1971) Decentralized PID controller settings Effect of noise & initial guess values are studied.
Optimization computational time considerably reduced from 1 hr and 20 mins 1 to 1 min and 26 s. 14 Closed-loop responses of the process & the identified FOPTD model for example 2. (Model & process response curves overlaps) . 15 15
Model IAE values ISE values Time (N.I) FOPTD Model identification results for without noise 11 Guess 1.2500 1.2500 0.1000 0.0177 1.1810 -4
Converged 2.5160 0.6608 0.4000 N.I: Number of iterations (in brackets); Time in Sec Table 1.2. Model identification results using decentralized PID controller for Example 2. 16 Example 1.3: Ogunnaike & Ray distillation column process (Ham & Kim, 1998)
Decentralized PID controller settings Effect of controller settings & initial guess values are studied.
Converged model parameters match well with actual FOPTD system.
Optimization computational time is 23 min and 20 s. 17 17 Closed-loop response of the process and the identified FOPTD models for example 1.3 18
Model IAE values ISE values 11 Guess 1.2739 10.0000 2.6000 1.7510 -3 6.3610 -8
Converged 0.8500 6.5995 0.8300 Table 1.3. FOPTD Model identification results for example 1.3 19 Guess values for SOPTD model parameters Design value of is taken as Diagonal process gains obtained from controller gain. The value of Time delay considered as same as closed time delay. Time constant obtained from settling time. Guess value for time constant is assumed as Approximated SOPTD Model Guess value for Damping coefficient 1 20 20 Objective Function for Optimization New method of excitation is proposed. The method, not only considered dynamics of step (up) change and also considered dynamics by removing step change(Combined step-up & step-down) Number of parameters to be estimated increases (from 12 to 16 parameters for TITO systems) (FOPTD to SOPTD) It is shown in this study that by the step response identification, closed-loop response of process and model are closely matching. But identified model parameters are not matching with that of the actual process 21 21 Simulation examples Identification Identification of SOPTD to SOPTD TITO system. In step response method, Identified model parameters not matching with actual SOPTD model Distinguish differences between identified models are shown by Nyquist & open-loop transient response plot. Identification of Higher order model to SOPTD TITO system. IAE & ISE values of identified model by step response method are higher compared to model identified by combined step-up & step-down method. Distinguish differences between identified models are shown by Nyquist plot Identification of SOPTD to SOPTD 3 3 system. Effect of different controller settings, noise & initial guess values are studied. Optimization computational time is 9 min and 14 s (max. computational Time)
22 Converged model parameters match well with actual SOPTD system for combined step-up & step-down response method
Effect of different controller settings, noise & initial guess values are studied.
Optimization computational time considerably reduced from 5 hr and 14 mins to 11 min and 11 s (max. computational Time) Example 2.1: Weischedel and McAvoy (1980) distillation column transfer function model Decentralized PID controller settings 23 Closed-loop step response of the process and the identified SOPTD model (from step-up (only) response) 24 24 Model IAE values ISE values
Table 2.1. Model identification results using PID controller for example 2.1. * step-up (only) response; ** Combined step-up and step-down response 25 Nyquist plot for actual system (solid line), identified SOPTD model: Combined step-up and step-down method (++), step-up (only) method (dot). . 25 Real axis 26 26 Closed-loop response of the process and the identified SOPTD transfer function model (Model & Process curves overlap) 27 27 Example 2.2: Niederlinski (1971) multivariable transfer function model Decentralized PID controller settings Low IAE & ISE values for SOPTD model identified by combined step-up & step-down excitation.
Effect of noise, controller parameters & initial guess values are studied.
Optimization computational time considerably reduced from 1 hr and 2 mins to 3 min and 3 s (max. computational Time). . 28 28 . Nyquist plot for actual system (solid line), identified SOPTD model: Combined step-up and step-down method (++), step-up (only) method (dot) (Example-2). Real axis 29 29 29 Model IAE values ISE values 11 Guess 1.2500 3.1250 0.1000 1.0000 Converged* 0.4999 0.3557 0.0498 0.7682 0.0132 5.67E-05 Converged** 0.5005 0.2818 0.1049 0.8734 0.0065 9.50E-06 21 Guess 2.0088 2.5000 0.1000 1.0000 Converged* 1.0008 0.2494 0.0250 0.9629 0.0145 5.85E-05 Converged** 1.0002 0.2061 0.0859 1.0254 0.0118 4.84E-05 12 Guess -2.0088 2.5000 0.1000 1.0000 Converged* -1.0015 0.1826 0.0250 1.2708 0.0112 8.79E-05 Converged** -0.9992 0.2267 0.0667 0.9727 0.0038 2.24E-06 22 Guess 3.3333 2.5000 0.1000 1.0000 Converged* 2.3959 0.4415 0.2020 0.9183 0.0165 8.46E-05 Converged** 2.4015 0.4110 0.1905 0.9901 0.0107 2.93E-05 Table 2.2. Model identification results using PID controller for example 2.2 * step-up (only) response; ** Combined step-up and step-down response 30 30 Closed-loop response of process and identified SOPTD transfer function model (Model & Process curves overlap) 31 31 31 Example 2.3: Transfer function considered by Garcia et al (2005). Decentralized PI controller Converged model parameters match well with actual SOPTD system for combined step-up & step-down response method
Effect of different controller settings, noise & initial guess values are studied.
Optimization computational time is 9 min and 14 s (max. computational Time) 32 32 32 Closed-loop response of the process and the identified SOPTD transfer function model (Model & Process curves overlap) 33 Independent design of PI/PID controller based on the equivalent transfer function (ETF) of individual loops with concept of decoupler.
Decomposing of multi-loop systems in to independent loops (Multi-single loop system) by using decouplers and then resulting model is approximated using the concept of ETF.
Open-loop responses of ETFs & EOTF models are compared.
Independent design of PI/PID controller was designed for each singles loop based on the corresponding ETFs.
Multivariable controller design based on ETF models 34
TITO system input-output relationship (Seborg et al.,2009)is
When , overall closed-loop transfer function model b/w and This can be rewritten as Perfect control approximation (Vu & Lee, 2010) 35 TITO decoupled control system Relationship between the input vector and process output vector (Bequette, 2003) 36 36 = Dynamic decouplers (Bequette, 2003 )expressed as Presence of decoupler, TITO system behaves like two independent loops and it is expressed as For TITO process 37 Eq. (22) & (27), both are equivalent. Eq. (22) is derived based on perfect control approximation. Eq. (27) is derived by introducing decoupler in to the system. 37 Wang et al.
have suggested a method to obtain the realizable decoupler where If the extra time delay is incorporated with the decoupler matrix, it will also affect the corresponding ETF model. For example = Then the decoupled Process model as
38 ETF model parameters from RNGA-RARTA (He et al. 2009)concepts: Integrated Error( ) b/w step input ( ) and output ( ) defined as average residence time( ) :
For FOPDT system Normalized Gain RNGA RARTA 39 Equivalent transfer function (ETF) is the equivalent transfer function when all other loops are closed. assume that ETFs have same structures as corresponding open loop transfer functions. For FOPDT model, the ETF is expressed (He et al., 2009)as SIMC (Skogestad, 2003)PI controller tuning parameters PI controller form as Based on ETF the PI controller is designed 40 Example 3.1: The transfer function matrix of VL column given by
By using the RGA and RNGA concepts, the ETF model parameters are deduced by 41 The ETF model matrix is expressed as: The actual EOTF models are derived using equations (10) and (11), and are expressed as: Simplified decoupler matrix The decoupled process models are expressed as: ; 42 Closed-loop responses (sequential step changes in the set point) for the VL column 43 Performance indices Ideal Normalized Inverted Proposed IAE 2.1830 0.7186 2.437 2.2640 2.179 0.0111 1.933 0.0156 0.0049 2.487 0.436 3.6120 0.002 2.432 0.0027 2.115 Disturbance 0.9950 1.438 0.9573 2.3840 0.995 1.438 0.954 1.460 ISE 1.6910 0.0947 1.689 0.7161 1.693 1.5410 -5 1.448 3.4910 -5
2.510 -6 1.884 0.0249 2.6930 5.7110 -7 1.775 1.2110 -6 1.537 Disturbance 0.0621 0.2279 0.0591 0.4063 0.0621 0.2179 0.0531 0.1928 Performance indices Ideal Normalized Inverted Proposed IAE * 5.3935 8.7490 4.6241 4.0663 ISE * 3.6697 5.1230 3.4681 2.9850 Comparison of IAE and ISE values for different decoupling control schemes for VL column Performance indices for VL column *Sum of main and interaction responses 44 Example 3.2: ISP reactor system (Chien et al.,1999)
ETF model parameters are 45 ETF model as expressed as The actual EOTF models The decoupled process models are expressed as ,
46 Comparison of IAE values Method used Change in set point IAE values IAE *
Proposed Method 0.9674 0.0024 1.8948 0.0141 0.9109 Centralized controller 0.7270 0.4039 2.7090 0.5917 0.9864 Inverted Decoupler 1.0813 0.0001 1.9604 0.0014 0.8776 * Sum of main and interaction responses 47 Conclusions It is shown by simulation examples, the proposed method for FOPTD/SOPTD model guess values able to give the faster and guaranteed in convergence.
Problem notified in convergence of SOPTD model parameters by conventional step response identification method.
New method of excitation (combined step-up & step-down) is proposed for the SOPTD model identification of multivariable systems.
Si. No. Simulation Examples No. of iterations Computational Time 1.1. WB Distillation Column 28 4 min and 17 s 1.2. Niderlinski transfer function model 24 1 min and 26 s 1.3. OR distillation Column 42 33 min and 14 s 48 48 It is shown the proposed method is robust for controller settings, noise and guess values perturbations (up to 10%).
Decomposing of TITO systems in to independent loops (Two-single loop system) by using dynamic decouplers and the resulting model is approximated using the concept of ETF.
PI Controller designed based on the ETF model using SIMC method. The control system performances are compared for the Ideal, Normalized, Inverted decoupling and centralized control system. Si. No. Simulation Examples No. of iterations Computational Time 2.1. WM Distillation Column 44 11 min and 11 s 2.2. Niderlinski transfer function model 34 3min and 3 s 2.3. Garcia et al. transfer function model 17 3 min and 17 s 49 (1)Viswanathan, P. K.; Toh, W. K.; Rangaiah, G. P. Closed-Loop Identification of TITO Processes Using Time-Domain Curve Fitting and Genetic Algorithms, Ind. Eng. Chem. Res. 2001, 40, 2818-2826. (2)Promod, S.; Chidambaram, M. Closed Loop Identification by Optimization Method for Tuning PID Controllers, Indian Chem. Engr., Section A, 2001, 43, No.2, 90-94. (3)Weischedel, K.; McAvoy, T. J. Feasibility of Decoupling in Conventionally Controlled Distillation Columns, Ind. Eng. Chem. Fundam., 1980, 19, 379-384. (4)Codrons, B. Process Modeling for Control: A Unified Framework Using Standard Black-Box Techniques, Springer-Verlag London Limited, 2005, p 48. (5)Niderlinski, A. A Heuristic Approach to the Design of Linear Multivariable Interacting Control Systems, Automatica, 1971, 7, 691-701. 50 References (6) Wood, R. K.; Berry, M. W. Terminal Composition Control of a Binary Distillation Column, Chemical Engineering Science, 1973, 28, 1707-1717. (7) Ham, T. W.; Kim, Y. H. Process Identification and PID Controller Tuning in Multivariable Systems, Journal of Chemical Engineering of Japan, 1998, 31, No.6, 941-949. (8) Vu, T. N. L.; Lee, M. Independent Design of Multi-loop Controllers for Interacting Multivariable Processes, J. Process Control, 2010, 922-933. (9) He, M. J.; Cai, W. J.; Ni, W.; Xie, L. H. RNGA Based Control System Configuration for Multivariable Processes, J. Process Control, 2009, 19, 1036- 1042. (10) Garcia, D.; Karimi, A.; Longchamp, R. PID Controller Design for Multivariable Systems Using Gershgorin Bands, IFAC World Cong., 2005. PID control in the Third Millennium (Lessons learned & the new approaches) R,Vilanova & A.Visioli (Ed.), Springer 2012
PID Control: New Identification & Design Methods Johnson ,M.A., Moradi, M.H. (Ed.), Springer, 2005
Handbook of PI & PID Controller Tuning Rules, 3 rd , ODwyer, A., (Ed.) Imperial College Press, 2009
Industrial Process Identification and Control Design, Liu,T. & F.Gao, Springer, 2012
Practical PID Control, A.Visioli, Springer, 2006
Advanced Process Identification and Control, E.Ikonen & K.Najim Marcel Dekker, New York, 2002.
Computer Control of Processes M.Chidambaram Narosa Publishers, New Delhi, 2002.
Advanced Practical Process Control B.Riffel & B.H.L.Betlem, Sprnger, 2004
53 System Identification; Theory for the user L.Ljung, Prentice-Hall, New Jersey, 1999
Multivariable System identification for Process control, Zhu, Y., Elsevier Science & Technology Books, 2001
System identification, Keesman, K.J., Springer, London, 2011
System identification and control Landau,I.D., Prentice-Hall, 1990.
C.Rajapandiyan and M.Chidambaram Closed loop identification of multivariable systems by optimization method Industrial & Engg Chemistry Research, V51,1324-1336 (2012)
C.Rajapandiyan and M.Chidambaram Closed loop identification of second order plus time delay (SOPTD) model of multivariable systems by optimization method Industrial & Engg Chemistry Research, V51, 9620-9633 (2012)
C.Rajapandiyan and M.Chidambaram Controller design for MIMO processes based on simple decoupled transfer functions and simplified decoupler Industrial & Engineering Chemistry Research, V51, 12398- 12410 (2012) 55 Decentralized control system of a TITO process. 56 56 Closed-loop response of the process and the identified FOPTD transfer function model for example 1 (with noise) using PID controller settings. 4 57 57 Parameters
Vu et al. 5 (0.7, 17.27; -0.13, 16.08) Loh et al. 6 (0.868, 3.246; -0.0868, 10.4) Viswanathan et al. 1 (0.2, 4.0, 3.0; -0.03, 2.0, 3.0) N. I. 16 21 28 Computational time (sec) 94 132 257 IAE * values 0.0229 0.0046 0.0066 ISE * values 6.5110 -6 3.3510 -7 4.4010 -7
Controller parameters Variation in guess values N. I IAE * values ISE * values Computational Time (Sec) Ham & Kim 4 +10% 25 0.004 1.8810 -7 152 -10% 25 0.0047 1.6210 -7 154 Vu et al. 5 (0.7, 17.27; -0.13, 16.08) +10% 14 0.0120 1.8410 -6 86 -10% 16 0.0060 4.0710 -7 94 Loh et al. 6 (0.868, 3.246; -0.0868, 10.4) +10% 26 0.0042 3.3110 -7 169 -10% 28 0.0045 3.4610 -7 172 Viswanathan et al. 1
(0.2, 4.0, 3.0; -0.03, 2.0, 3.0) +10% 37 0.0066 4.3710 -7 352 -10% 30 0.0066 4.4010 -7 313 Effect of changing controller settings on number of iterations and computational time for example 1. Effect of initial guess values on number of iterations and computational time for example 1. * Sum of main and interaction response values, N.I: Number of iterations 58 58 Closed-loop responses of the corrupted process and the identified FOPTD transfer function model for example 1.2 (with noise). 59 59 Parameters N. I IAE * values ISE * values Computational Time (min) Ham & Kim (1998)
(2.794, 4.230, 1.102; -0.524, 5.635, 1.110; 3.835, 6.006, 1.069) 50 0.6222 0.0096 30.53 Chien et al (1999) (1.08, 4.25; -0.2333, 3.32; 2.78, 5.24) 42 0.056 4.08E-05 33.23 Controller parameters Variation in guess values N. I IAE * values ISE * values Computational Time (min) M.Lee -10% 25 0.0939 4.32E-04 20 +10% 27 0.0476 2.37E-05 21.31 Ham & Kim 4 (2.794, 4.230, 1.102; - 0.524, 5.635, 1.110; 3.835, 6.006, 1.069) +10% 49 0.7528 1.4110 -2 30.17 -10% 32 0.7468 1.3010 -2 21.23 Chien et al. (1999) (1.08, 4.25; -0.2333, 3.32;2.78, 5.24) +10% 48 0.0562 4.6810 -5 38.48 -10% 41 0.0295 9.8610 -6 32.96 Effect of initial guess values on the number of iterations and computational time for example 1.3. Effect of changing controller settings on number of iterations and computational time for example 1.3. 60 60 Example 4: Industrial-Scale Polymerization reactor Example 5: Benzene-toluene distillation column 61 61 Observations from Viswanathan et al. method: Computational time for optimization ranges from 50 min to 3 hrs (for different controller settings & test timings). Converged values from GA, are not reliable and these values are taken as initial guess for BFGS method. After BFGS also, global minimum not guaranteed. General method not provided to get the upper & lower bounds. If they obtain the lower IAE values also, not able to get the exact model parameters. 62 Genetic Algorithms: stochastic global search method (mimics the process of natural evolution). Starting at several independent points and searching in parallel. Very big population size consumes more time for finding optimum solution. Parameters used in the floating point GA Populations size 200 Number of generations 100 Probability of crossover 0.6 Probability of mutation 0.2 Tournament Size 5.0 63 1. Generate an initial, random population of individuals.
2. Evaluate their fitness.
3. Select the fittest members of the population. Reproduce using a probabilistic method. Implement crossover operation on the reproduced chromosomes. Execute mutation operation with low probability.
4. Repeat step 2 until a predefined convergence criterion is met. 64 No absolute assurance, GA will find global optimum.
GA cannot assure constant optimization times.
Real time implementations are limited because of random solutions and convergence.
Converge very slow, especially near an optimum.
Determining of termination criterion is not straight forward.
Insufficiently converged solutions can produce misleading. Disadvantages of GA 65 n th order system From the Bode stability criteria, Phase angle Amplitude Ratio 66 66 Codrons reported that for same controller different models shows same closed loop behavior even though they have different open-loop behavior. Two examples are considered, Integrating system First order lag system Proportional controller 67 67 67 67 Open-loop response of process and identified SOPTD transfer function models of Example 2.1(Model & Process curves overlap) 68 68 Model IAE values ISE values 11 Guess 0.2500 5.0000 1.0000 1 . 0000 Converged 0.5948 3.8334 0.7693 1.0948 0.3386 2.90E-03 21 Guess 0.2152 5.0000 1.3000 1.0000 Converged 0.3673 1.2488 1.5696 2.2692 0.2057 8.86E-04 12 Guess -0.1824 5.0000 1.0000 1.0000 Converged -0.4653 4.3886 0.7786 0.9779 0.4083 3.40E-03 22 Guess -0.1667 5.0000 1.0000 1.0000 Converged -0.4988 1.7106 0.7267 1.5988 0.3889 5.70E-03 Model identification results using PID controller with measurement noise Controller PID Controller PI controller Perturbation in Guess values -10% 0% +10% -10% 0% +10% IAE * values 0.0477 0.0618 0.0618 0.0519 0.0643 0.0643 ISE * values 3.49E-05 5.03E-05 5.03E-05 2.64E-05 4.35E-05 4.35E-05 Number of Iterations 33 41 40 33 37 44 Computational Time (Sec) 323 390 379 509 567 671 Effect of initial guess values for example 2.1 69 69 69 Mode of Controller PI b Controller PI a controller Perturbation in Guess values -10% 0% +10% -10% 0% +10% IAE * values 0.0302 0.0327 0.0282 0.0463 0.0275 0.0347 ISE * values 9.39E-05 8.94E-05 8.25E-05 1.97E-04 4.52E-05 7.67E-05 Number of Iterations 30 28 34 37 46 42 Computational Time (Sec) 158 151 183 130 143 132 Effect of initial guess values for example 2.2 70 PID control in the Third Millennium (Lessons learned & the new approaches) R,Vilanova & A.Visioli (Ed.), Springer 2012
PID Control: New Identification & Design Methods ( Johnson ,M.A., Moradi, M.H. (Ed.), Springer, 2005
Handbook of PI & PID Controller Tuning Rules, 3 rd , ODwyer, A., (Ed.) Imperial College Press, 2009
Industrial Process Identification and Control Design, Liu,T. & F.Gao, Springer, 2012
Practical PID Control, A.Visioli, Springer, 2006
Advanced Process Identification and Control, E.Ikonen & K.Najim Marcel Dekker, New York, 2002.
Computer Control of Processes M.Chidambaram Narosa Publishers, New Delhi, 2002.
Advanced Practical Process Control B.Riffel & B.H.L.Betlem, Sprnger, 2004
System Identification; Theory for the user L.Ljung, Prentice-Hall, New Jersey, 1999