Professional Documents
Culture Documents
pubs.acs.org/IECR
ABSTRACT: A simulation-based optimization framework is introduced. First, a base case design is conducted using the
commercial simulator. Then, design variables are decided. Next, minimum and maximum ranges of the design variable are
determined. We will call this the design space. A comprehensive simulation of the design space is executed. Next, empirical
modeling of this design space is performed. This is called the process mapping step. After verification of the model using a test
data set, the optimization problem is solved using the developed data-driven model. The methodology is applied to the optimal
design of a natural gas liquefaction process for an offshore liquefied natural gas (LNG) plant.
© 2014 American Chemical Society 5539 dx.doi.org/10.1021/ie403507p | Ind. Eng. Chem. Res. 2014, 53, 5539−5544
Industrial & Engineering Chemistry Research Article
inequality form. In our framework, the optimization space is heat exchange. ASPEN HYSYS v7.3 is used for steady-state
confined to the design space. modeling usingthe Peng−Robinson property package.19
Step 6. Split the design space into appropriate discrete points Natural gas at 50 bar is cooled to −20 °C in LNG-100 and
and run a comprehensive simulation of the total range. This is then liquefied in LNG-101 with a discharge temperature of −90
the “data extraction” step. °C. The LNG-101 module is where the phase change of natural
Step 7. Construct a new model of the design space using gas to LNG occurrs. LNG is further cooled to −110 °C and
empirical modeling. This is called the “process mapping” step. −155 °C at LNG-102 and LNG-103, respectively. LNG is
The developed empirical model is the short-cut data-driven expanded in the Joule−Thompson valve (JT1) to 1.2 bar
model of the target process for the design space only. It should resulting in a temperature decrease to −165.2 °C.
not be used for extrapolation. The input variables of the model There is no phase change in the nitrogen refrigerant, and the
are the design variables. The model is to predict the objective cycle is operated only in vapor phase. The nitrogen gas is
function value f(X) and inequality constraint function value
compressed to 55 bar by compressor modules C1, C2, C3, and
g(X). Predicting multivariate outputs is an important feature
C4. Nitrogen after the second stage of compressor (C2) is split
because the values of inequality constraints should be included
in addition to the estimation of the objective function value. in TEE-100 to streams 6 and 8 to be compressed further in C3
Step 8. Validate the empirical model using the test data set. If and C4, respectively. HX1, HX2, and HX3 are air coolers to
the empirical model shows reasonable prediction performance, cool down the compressor discharge streams to 30 °C. The
then it can be used for optimization. refrigerant at 10 °C (stream 12) is precooled with the natural
Step 9. Solve the optimization problem of the process using gas feed stream (NG2) to −20 °C in LNG-100. Then, the
the developed empirical model. refrigerant is split in the TEE-101 module to streams 14 and 16.
Stream 14 is expanded to 9 bar at expander X1 and is mixed
3. BASE CASE DESIGN OF TURBO-EXPANDER with stream 19 to form stream 20, a cold stream that liquefies
PROCESS the natural gas in LNG-102. Nitrogen in stream 16 is further
cooled at LNG-101 to −90 °C and later expanded to 9 bar at
3.1. Step 1: Base Case Design. Figure 2 shows the process expander X2. This cold stream at −159.8 °C is used as the
flow diagram of the natural gas liquefaction process using
coldest stream to ensure that the temperature of LNG falls to
nitrogen as the refrigerant. The process is based on a United
−155 °C. Turbo expander X1 is arranged to drive compressor
States patent18 and in-house data for 1 MTPA (million ton per
annum) LNG production. The natural gas feed stream consists C3, and turbo expander X2 is arranged to drive compressor C4.
of 5.7 mol % of nitrogen, 94.1 mol % of methane, and 0.2 mol So, power produced from the expanders is recovered in the
% of ethane. Streams 1 to 22 are nitrogen streams (colored blue compressors. The compressors need not be separate units and
in Figure 2) forming a cycle, while NG1 to NG7 are natural gas can be connected to a common shaft.
streams (colored green in Figure 2). Natural gas and nitrogen 3.2. Step 2: Specifying Design Variables. A total of six
inlet temperature values are set to 10 °C using coolers HX4 and design variables are to be manipulated in the target process.
HX5, respectively. The double expander process is applied to The design variables include the refrigerant mass flow rate (or
liquefy the natural gas at 50 bar to −155 °C (stream NG6). refrigerant to natural gas ratio), split ratio of TEE-101 (flow
Four LNG heat exchanger modules (LNG-100, LNG-101, ratio of stream 14), discharge pressure of the refrigerant at
LNG-102, and LNG-103) are used to simulate multistream expander (streams 15 or 18), and discharge temperature values
5541 dx.doi.org/10.1021/ie403507p | Ind. Eng. Chem. Res. 2014, 53, 5539−5544
Industrial & Engineering Chemistry Research Article
of natural gas at LNG exchangers LNG-100, LNG-101, and multiple effects on the process and should be determined by
LNG-102 (streams NG3, NG4, and NG5). optimization.
3.3. Step 3: Optimization Formulation. The optimiza- 4.2. Step 6: Comprehensive Simulation of Design
tion formulation is as follows Space. The six design variables have lower and upper bounds.
The range for each variable is divided into six intervals. A
min Z (total shaft work) (5)
comprehensive simulation, that is six to the sixth power, is
s.t. hI(XD) = 0 performed for data extraction. An external operator such as
(6)
MATLAB is used for easy handling of data input and output.
hE(XD) = 0 (7) The data collected are the objective function value and
minimum temperature approach values for each LNG ex-
MTA k ≥ 3 (for k changer. Therefore, a data table of six input variables and five
output variables is constructed through comprehensive
= LNG‐100, LNG‐101, LNG‐102, and LNG‐103) simulation.
(8) 4.3. Step 7: Process Mapping of Design Space Using
where XD is a vector of design variables specified in Section 3.2. Empirical Modeling. Empirical modeling is used for process
The objective function for optimal design of the liquefaction mapping. Examples of empirical modeling include multiple
process in eq 5 is to minimize the total shaft work, which means linear regressions, partial least-squares regressions,20 and other
compression work of C1 and C2. The equality constraints in eq useful machine learning methods. An artificial neural network is
6 include all the thermodynamic equations together with unit useful when estimating various output variables from various
operation equations. Additional equality constraints to be input variables.21 ANN is a well-known technique to model
expressed explicitly in eq 7 will be explained in Section 3.4. The nonlinear characteristics applied to a variety of chemical
inequality constraints in eq 8 represent the minimum engineering field.22
temperature approach (MTA) of each LNG exchanger. Here, In this study, a feed-forward back-propagation network with
the lower limit is 3 K. one input layer−one hidden layer−one output layer config-
3.4. Step 4: Providing Additional Constraints in the uration is employed. The input layer consists of six nodes, each
Simulator. The temperature value of stream 13 is set to be corresponding to the six design variables. The hidden layer has
equal to the temperature of stream NG3. In addition, the 20 nodes. The output layer has five nodes, one for the objective
temperature values of stream 17 and NG4 are set to be equal. function and four for the minimum temperature approach
The pressure values of streams 15 and 18 are also set to be values for LNG exchangers. A tan-sigmoid function in eq 9 is
equal. Pressure values of streams 7 and 9 are also set to be equal used for the activation function.
and are fixed at 55 bar. The split ratio of TEE-100 is not a exp(x) − exp( −x)
design variable and is rather a variable that is adjusted y=
exp(x) + exp( −x) (9)
depending on split ratio of TEE-101 and expander discharge
pressure value. Discharge pressure of C2 (stream 4) is also a
variable that is adjusted depending on the power recovered in 5. RESULTS AND DISCUSSION
C3 and C4. These design specifications are easily modeled in 5.1. Step 8: Empirical Modeling Validation. A test data
process simulators. set of 5000 randomly sampled observations is utilized for
4. SIMULATION-BASED OPTIMIZATION Table 2. Prediction Performance of the Neural Network
4.1. Step 5: Determining Design Space. Table 1 shows Model
the design variables, their initial points, and variable range. The
design space is the operating window that we are interested in. output variable average error RMSE dimension of RMSE
To minimize the shaft work of compressor, nitrogen flow total shaft work 0.029% 3.611 [kW]
rate (or refrigerant to natural gas ratio) should be decreased MTA of LNG-100 0.017% 0.006 [K]
from the initial point. Split ratio of TEE-101 (flow ratio of MTA of LNG-101 0.014% 0.043 [K]
stream 14 to stream 16) should be increased from the initial MTA of LNG-102 0.042% 0.011 [K]
point to minimize the heat duty of LNG-101, which makes it MTA of LNG-103 0.127% 0.049 [K]
possible to further decrease the refrigerant requirement. The
expander discharge pressure value should be increased from the empirical model validation. The model predicts five output
initial point because the compressor discharge pressure is fixed variables, total shaft work, and minimum temperature approach
in this case, and the lower expander discharge pressure results values for each LNG exchanger. Validation of the model means
in higher compression ratio and higher power consumption. comparing the predicted values with the original simulator data.
The outlet temperature value of each LNG exchanger has The results show excellent agreement between the simulator
5542 dx.doi.org/10.1021/ie403507p | Ind. Eng. Chem. Res. 2014, 53, 5539−5544
Industrial & Engineering Chemistry Research Article
Table 4. Empirical Model and Simulator Values Using Optimized Design Variables
category variable empirical model simulator dimension
design variable N2 refrigerant flow rate 1078 1078 [ton/h]
split ratio of TEE-101 0.5900 0.5900 −
expander discharge pressure 9.6483 9.6483 [bar]
LNG-100 outlet temperature −20.00 −20.00 [°C]
LNG-101 outlet temperature −85.96 −85.96 [°C]
LNG-102 outlet temperature −110.00 −110.00 [°C]
objective function total shaft work 45938 45941 [kW]
constraints LNG-100 MTA 3.00 3.01 [K]
LNG-101 MTA 3.00 2.98 [K]
LNG-102 MTA 4.71 4.71 [K]
LNG-103 MTA 3.00 3.00 [K]
and data-driven model. Average error percent and root mean apply many design specifications into the simulator that need
squared error (RMSE) of the values are tabulated in Table 2. some iteration steps to solve. The main disadvantage of the
The average error is less than 0.2%, and the difference between suggested methodology is that the data-extraction step can be
simulator and neural network model is negligible. Therefore, time consuming, especially with many design variables.
the developed empirical model can be utilized as a Development of the empirical model with the least data set
representative model for the target design space instead of available will be discussed in future work.
the commercial simulator itself.
It has been shown that empirical modeling of the design
6. CONCLUSION
space shows great performance of process mapping. However,
the prediction performance may vary with the dimension of Commercial process simulators are effective when modeling
design space, broadness of each input variable range, structure complex processes such as a liquefaction process with
of the empirical model, etc. Process engineers should predefine multistream heat exchangers and recycle streams. For the
a tolerance for the prediction error and only use the proposed optimal design of such a process using the simulator
methodology when the performance of the developed data- environment, this paper introduced a simulation-based
driven model is acceptable. optimization framework using a systematic process mapping
5.2. Step 9: Optimization. For the initial design point, technology. The main concept is to develop a short-cut data-
optimization using MATLAB linked with HYSYS does not driven model. The method includes a comprehensive data
converge to a value and fails to solve the problem. Case studies extraction phase of the process in a prespecified operating
with different initial points were also performed, and the results range. The design space of the process is modeled using an
of conventional optimization showed a high dependence on
empirical modeling technique such as an artificial neural
initial value. However, when the process design space is
network. The performance of the proposed method was shown
mapped using the suggested technique, the optimization
problem is solved regardless of the starting point. The by use of a case study of a natural gas liquefaction plant for
optimization results are tabulated in Table 3. floating production, storage, and off-loading. The proposed
To finally check the reliability of the empirical model, the optimization framework can be applied to any process design
optimized values need to be compared with the simulator and real-time optimization.
values. Table 4 shows the results, and the difference between
the empirical model and simulator is in the acceptable range.
The advantage of the proposed optimization framework is
■ AUTHOR INFORMATION
Corresponding Author
that process designers can take full advantage of the process
*Tel: +82-2-880-1887. E-mail: chhan@snu.ac.kr.
simulators. Process simulators are easy to handle and inherently
calculate unit operation equations with a vast and accurate Notes
thermodynamic database. Also, process designers can readily The authors declare no competing financial interest.
5543 dx.doi.org/10.1021/ie403507p | Ind. Eng. Chem. Res. 2014, 53, 5539−5544
Industrial & Engineering Chemistry Research
■
Article
■ REFERENCES
(1) Venkatarathnam, G. Cryogenic Mixed Refrigerant Processes;
Springer: New York, 2008.
(2) Gerdsmeyer, K.-D.; Isalski, W. H. On-Board Reliquefaction for
LNG Ships. In Proceedings of the Gas Processors Association Europe
Conference, London, 2005.
(3) Li, Q. Y.; Ju, Y. L. Design and analysis of liquefaction process for
offshore associated gas resources. Appl. Therm. Eng. 2010, 30, 2518−
2525.
(4) Finn, A. J. Floating LNG Plants-Scale-Up of Familiar Technology.
In Proceedings of the Gas Processors Association 88th Annual
Convention, San Antonio, TX, 2009.
(5) Foglietta J. H. U.S. Patent 6,412,302, 2002.
(6) Dubar, C. A. U.S. Patent 5,768,912, 1998.
(7) Wang, M.; Zhang, J.; Xu, Q.; Li, K. Thermodynamic-analysis-
based energy consumption minimization for natural gas liquefaction.
Ind. Eng. Chem. Res. 2011, 50, 12630−12640.
(8) Shah, N. M.; Hoadley, A. F. A targeting methodology for
multistage gas-phase auto refrigeration processes. Ind. Eng. Chem. Res.
2007, 46, 4497−4505.
(9) Cao, W. S.; Lu, X. S.; Lin, W. S.; Gu, A. Z. Parameter comparison
of two small-scale natural gas liquefaction processes in skid-mounted
packages. Appl. Therm. Eng. 2006, 26, 898−904.
(10) Song, K.; Jeong, C.; Nam, J.; Han, C. Hybrid compressor model
for optimal operation of compressed dry air system in LCD
production industry. Ind. Eng. Chem. Res. 2012, 51, 4998−5002.
(11) Singh, K. K.; Shenoy, K. T.; Mahendra, A. K.; Ghosh, S. K.
Artificial neural network based modelling of head and power
characteristics of pump-mixer. Chem. Eng. Sci. 2004, 59, 2937−2945.
(12) Fissore, D.; Barresi, A. A.; Manca, D. Modelling of methanol
synthesis in a network of forced unsteady-state ring reactors by
artificial neural networks for control purposes. Chem. Eng. Sci. 2004,
59, 4033−4041.
(13) Kadlec, P.; Gabrys, B.; Strandt, S. Data-driven soft sensors in the
process industry. Comput. Chem. Eng. 2009, 33, 795−814.
(14) Becker, T.; Enders, T.; Delgado, A. Dynamic neural networks as
a tool for the online optimization of industrial fermentation. Bioprocess
Biosyst. Eng. 2002, 24, 347−354.
(15) Nielsen, D.; Pitchumani, R. Intelligent model-based control of
preform permeation in liquid composite molding processes, with
online optimization. Composites, Part A 2001, 32, 1789−1803.
(16) Gonzaga, J. C. B.; Meleiro, L. A. C.; Kiang, C.; Filho, R. M.
ANN-based soft-sensor for real-time process monitoring and control
of an industrial polymerization process. Comput. Chem. Eng. 2009, 33,
43−49.
(17) Rotem, Y.; Wachs, A.; Lewin, D. R. Ethylene compressor
monitoring using model-based PCA. AIChE J. 2000, 46, 1825−1836.
(18) Dubar, C. A.; Tu, O. L. M. U.S. Patent 6,250,244, 2001.
(19) Peng, D. Y.; Robinson, D. B. A new two-constant equation of
state. Ind. Eng. Chem. Fundam. 1976, 15, 59−64.
(20) Geladi, P.; Kowalski, B. R. Partial least-squares regression: a
tutorial. Anal. Chim. Acta 1986, 185, 1−17.
(21) Bulsari, A. B. Neural Networks for Chemical Engineers; Elsevier
Science: Amsterdam, The Netherlands, 1995.