Computers and Chemical Engineering 31 (2007) 10551063

Optimization of dynamic systems: A trigonometric differential evolution approach

Rakesh Angira , Alladwar Santosh
Department of Chemical Engineering, Birla Institute of Technology & Science, Pilani 333 031, Rajasthan, India Received 30 January 2006; received in revised form 4 September 2006; accepted 22 September 2006 Available online 15 November 2006

Abstract In many chemical engineering applications, one frequently encounters dynamic optimization problems. The solution of these types of problems is usually very difcult due to their highly nonlinear, multidimensional and multimodal nature. In fact, several deterministic techniques have been proposed to solve these problems but difculties related to ease of implementation, global convergence, and good computational efciency have been frequently found. Recently, evolutionary algorithms (EAs) are gaining popularity for solving complex problems encountered in many engineering disciplines. They are found to be robust and more likely to locate global optimum as compared to several deterministic (gradient based) optimization methods. This paper deals with the application and evaluation of one such algorithm called trigonometric differential evolution (TDE) algorithm for solving dynamic optimization problems encountered in chemical engineering. This is a modied version of differential evolution (DE) which provides enhanced convergence speed. Both DE and TDE algorithms have been applied to seven dynamic optimization problems (ve optimal control problems and two kinetic parameter estimation problems) taken from recent literature. The obtained numerical simulation results indicate better performance of TDE as compared to that of DE particularly for problems involving large number of control stages. 2006 Elsevier Ltd. All rights reserved.
Keywords: Dynamic optimization; Optimal control; Kinetic parameter estimation; Chemical processes; Differential evolution; Trigonometric differential evolution

1. Introduction In many chemical engineering applications, one frequently encounters dynamic optimization problems. The solution of these types of problems is usually very difcult due to their highly nonlinear and multidimensional nature and also due to the presence of constraints on both state and control variables and implicit process discontinuities (Lim, Tayeb, Modak, & Bonte, 1986; Barton, Allgor, Feehery, & Galan, 1998). There has been a vigorous effort in the last 10 years to develop efcient algorithms for dynamic optimization. These algorithms can be divided into two classes, the stochastic approach and the deterministic approach. Most of the traditional algorithms based on gradient methods have the possibility of getting trapped at local optimum depending upon the degree of nonlinearity and initial guess but population based algorithms are found to have better global perspective than the traditional methods (Onwubolu &

Corresponding author. Tel.: +91 1596 245073x205/216/91 1596 245665. E-mail address: angira@bits-pilani.ac.in (R. Angira). URL: http://discovery.bits-pilani.ac.in/discipline/chemical/rangira/

Babu, 2004). Deterministic methods and their applications are detailed in Floudas (2000). During the last 5 years, some deterministic algorithms have been considered for dynamic optimization. The BB method (Esposito & Floudas, 2000), a deterministic spatial branch and bound global optimization algorithm (Papamichail & Adjiman, 2004), physically bounded GaussNewton method (Tang, Zhang, Linninger, Tranter, & Brezinsky, 2005) which might fail on certain problems with saddle points, and terrain methods (Lucia, DiMaggio, & Depa, 2004; Lucia, Bellows & Octavio, 2005) to name a few. The BB method and spatial branch and bound method, however guarantee of attaining the global solution, but require rigorous values for the parameters needed or rigorous bounds on these parameters and therefore are not easy to implement. The terrain method overcomes the saddle point limitation of physically bounded GaussNewton method and provides a reliable means of global optimization of differential equation models (Lucia et al., 2005). Several researchers have used stochastic optimization to solve dynamic optimization problem. The iterative dynamic programming (Luus, 1994; Dadebo & Mcauley, 1995), integrated

0098-1354/$ see front matter 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.compchemeng.2006.09.015

1056

R. Angira, A. Santosh / Computers and Chemical Engineering 31 (2007) 10551063

controlled random search for dynamic systems (ICRS/DS) and adaptive randomly directed search for dynamic systems (ARDS/DS) (Carrasco & Banga, 1997), direct search procedure of Luus and Jaakola (1973) (Luus & Hennessy, 1999), evolutionary algorithms like genetic algorithms (GA), differential evolution and ant colony algorithm (Wang & Chiou, 1997; Pham, 1998; Chiou & Wang, 1999; Rajesh, Gupta, Kusumakar, Jayaraman, & Kulkarni, 2001; Sarkar & Modak, 2003; Upreti, 2004; Angira & Babu, 2005; Angira & Alladwar, 2005; Angira, 2005; Babu & Angira, 2006), etc. Differential evolution (Storn & Price, 1995) is an efcient, effective and robust evolutionary optimization technique and is more likely to obtain the global solution as compared to several traditional optimization methods (Storn & Price, 1996; Corne, Dorigo, & Glover, 1999; Fan & Lampinen, 2003; Angira & Babu, 2003; Angira, 2005; Angira & Babu, 2006; Babu & Angira, 2006, etc. to name a few). Recently, Fan and Lampinen (2003) proposed trigonometric differential evolution (TDE) algorithm where a trigonometric mutation operation is embedded into DE algorithm to enhance its convergence velocity. And to adjust the balance between the convergence rate and robustness, a new parameter Mt , is introduced. Fan and Lampinen (2003) found convergence rate of TDE to be higher than that of the original DE. So far, TDE algorithm has not been applied and evaluated for solving dynamic optimization problems except Angira and Alladwar (2005). This paper deals with the application and evaluation of TDE algorithm for solving dynamic optimization problems encountered in chemical engineering. Both DE & TDE algorithms are used to solve ve nonlinear, multidimensional and multimodal optimal control problems and two kinetic parameter estimation problems taken from recent literature. Also, the performance of TDE algorithm is compared with that of original DE algorithm and other methods by which the particular problem has been solved in literature. This paper is organized as follows. The following section briey discusses the TDE algorithm. Simulation results obtained using DE and TDE are detailed in Section 3. Finally concluding remarks are made in Section 4. 2. TDE in brief In TDE algorithm (Fan & Lampinen, 2003), the trigonometric mutation operation is embedded into the original DE/rand/1/binversion to enhance the convergence rate. Therefore, the main difference between DE and TDE is the way the mutation operation is performed (i.e. the way the noisy random vector is generated). The mutation operations of DE and TDE are explained below: 2.1. Mutation operation in DE Select three distinct vectors/population points X[a][j], X[b][j], and X[c][j] randomly from the current population (primary array) other than the target vector, i.e. X[i][j] and create noisy random vector (NRV) with a probability CR as shown below: NRV = X[c][j ] + F (X[a][j ] X[b][j ]); (1)

where a, b, and c are randomly selected population points such that they satisfy: a = b = c = i; i = 1 to NP (population size), and j = 1 to D (number of decision variable). The noisy random vector and the target vector are then subject to crossover operation to create trial vector. Fitness of trial and target vectors is compared and the one with better tness will survive to next generation. Thus, the mutation operation of DE makes use of the vector differentials between the existing population members for determining both the degree and direction of perturbation. 2.2. Trigonometric mutation operation The trigonometric mutation operation is performed according to the following formulation: NRV = Xavg [j ] + F1 (X[a][j ] X[b][j ]) + F2 (X[b][j ] X[c][j ]) + F3 (X[c][j ] X[a][j ]); (2) where Xavg [j] = (X[a][j] + X[b][j] + X[c][j]/3, F1 = (pb pa ), F2 = (pc pb ), F3 = (pa pc ), pa = |f(X[a][j])|/P, pb = |f(X [b][j])|/P, pc = |f(X[c][j])|/P, and P = |f(X[a][j])| + |f(X[b][j])| + |f(X[c][j])|. As it can be seen from the above formulation, the vector to be perturbed is taken to be the average of three randomly selected vectors (a, b, and c). The perturbation to be imposed to this vector is then made up with a sum of three weighted vector differentials. F1 , F2 , and F3 , are the weight terms applied to these vector differentials. It is noted that the trigonometric mutation is a rather greedy operator since it biases the Xavg [j] strongly in the direction where the best one of three individuals chosen for the mutation is lying. The reader is referred to Fan and Lampinen (2003) for a further description of the trigonometric mutation operation. 2.3. TDE algorithm The TDE algorithm which include both the original DE mutation scheme and the trigonometric mutation scheme is outlined as follows: (1) Input system data and choose the population size (NP), the scaling factor (F), the crossover constant (CR), and mutation probability (Mt ). (2) Generate initial population in upper and lower bounds of each decision variables. (3) Evaluate the each individual/vector in population. A differential equation solver must be used in this step to obtain the dynamic state variables and the steady-state variables. (4) Perform mutation, crossover operation to obtain the offspring. (4.1) Perform mutation operation (4.1.1) Perform trigonometric mutation using Eq. (2) with a probability Mt . (4.1.2) Perform original DEs mutation using Eq. (1) with a probability (1 Mt ). (4.2) Perform crossover operation.

R. Angira, A. Santosh / Computers and Chemical Engineering 31 (2007) 10551063

1057

(5) Check whether the parameters in the trial vector are restricted within the bound. If an individual of this new population is outside the bounds, then this individual is replaced by randomly generating between upper and lower bounds. (6) Evaluate the offspring as in step 3. (7) Select the best between trial and target vectors. (8) Repeat step (4) to (6) until the termination criterion is satised. In the TDE algorithm, the parameter Mt provides a convenient way to keep a good balance between fast convergence and global optimum searching capability. It is to be noted that for a special case, Mt = 0, TDE effectively reduces into original DE. 3. Simulation results and discussion Both DE and TDE algorithms are applied to solve ve optimal control problems and two kinetic parameter estimation problems taken from recent literature. The problems are chosen in a manner to illustrate the ability of both the algorithms to cater the problems of varying levels of difculty. Several simulations are done to tune the Mt parameter of TDE algorithm for each problem and corresponding tuned value is reported. For each problem, 50 runs of each DE and TDE algorithms are carried out in order to ensure that the seed used for the random number generator did not bear any inuence on the quality of the results obtained. Both algorithms are coded in C language (Microsoft Visual C++ 6.0, compiler). The reported results of this study are obtained using an IBM computer (Pentium-IV/2.40 GHz/RAM 256 MB). 3.1. Optimal control problems For the numerical solution of the optimal control problems considered in the present study, only the control parameters are discretized into D control stages. And the dynamic system is then integrated in each control stage to evaluate the objective function and the constraints. To solve the differential equations of the process model, the RungeKutta fourth-order method is used for rst three problems and for the other two problems (i.e. bifunctional catalyst and the fed batch bioreactor problem), the fth-order RungeKutta Fehlberg method with CashKarp parameters, and adaptive step-size control (Press, Teukolsky, Vetterling, & Flannery, 2002) is used. 3.1.1. Plug ow reactor parallel reaction problem In this problem, an optimal control of plug ow reactor where the parallel reaction AB and AC occurs simultaneously
Table 1 Comparison of results for plug ow reactor parallel reaction problem S. no. No. of control stages (D) Best objective function value (J) Using TDE 1 2 3 4 10 20 40 80 0.572242 0.573297 0.573478 0.573527 Using DE 0.572242 0.573297 0.573480 0.573527 From literature 0.572241 0.573490 0.573480 0.573530 Average CPU-time (s) TDE 0.33718 0.33468 0.81812 5.391 DE 0.10532 0.4072 2.0328 26.32874
k1 k2

Fig. 1. Fractional difference variation for plug ow reactor parallel reaction problem (80 control stages).

with rate constant k1 and k2 is considered. The goal is to maintain the value of control variable u (which is a function of rate constants, reactor length, and ow velocity) along the length of reactor such that the yield of component B at exit of reactor is maximized. The dimensionless model describing the system dynamics of the reactant and product concentrations is given in literature (Logsdon & Biegler, 1989; Dadebo & Mcauley, 1995). The control variable is discretized into D number of intervals or control stages as reported in (Dadebo & Mcauley, 1995; Angira & Alladwar, 2005). The key parameters used are F = 0.5, CR = 0.8, NP = 2D and Mt = 0.32. The results of 50 independent runs of both the algorithms for eighty control stages are shown in Fig. 1. The fractional difference (f) is given by f = 1 J/Jbest , where Jbest is the best objective function value reported in literature (Dadebo & Mcauley, 1995) and J is obtained objective function value. The overall accuracy of 50-f values is quantied by their low average of 0.000006 in the range 0.0000050.000007 for DE and 0.000006 in the range 0.0000050.000006 for TDE. The precision of results is quantied by standard deviation of zero for both DE and TDE algorithms. Table 1 shows the comparison of results for both the algorithms using different number of control stages. CPU-time is the average of 50 different runs. From Table 1, it is clear that for all control stages the CPU-time required to obtain global optimum using TDE is less than that using DE algorithm except for ten control stages. TDE is approximately two times for forty stages

1058

R. Angira, A. Santosh / Computers and Chemical Engineering 31 (2007) 10551063

Fig. 2. Fractional difference variation for batch reactor consecutive reaction problem (80 control stages).

Fig. 3. Fractional difference variation for plug ow reactor catalyst blend problem (40 control stages).

and ve times for eighty stages faster than DE. The control prole obtained using DE and TDE is same and matches with that reported in Dadebo and Mcauley (1995). 3.1.2. Batch reactor consecutive reaction problem In this problem, the objective is to obtain the optimal temperature prole that maximizes the yield of the intermediate product B for a xed batch time in a batch reactor where the reaction ABC occurs. The model equations describing the system dynamics of the reactant and product concentrations are given in Dadebo and Mcauley (1995). The control variable is discretized into number of intervals or control stages (D) as reported in (Dadebo & Mcauley, 1995; Angira & Alladwar, 2005). The key parameters used are F = 0.5, CR = 0.8, NP = 2D and Mt = 0.42. The results of 50 different runs for each algorithm with 80 control stages are shown in Fig. 2. The overall accuracy of 50-f values is quantied by their low average of zero for DE and 1.188 1006 in the range 0.0000010.000005 for TDE. The precision of results is quantied by standard deviation of zero for DE and 7.18275 1007 for TDE. Table 2 shows the comparison of results for both the algorithms for different number of control stages. CPU-time is the average of 50 different runs. From Table 2, it is clear that for all control stages the average CPU-time required to obtain global optimum using TDE is less than that using DE algorithm. TDE is approximately six times for fty stages and nine times for 80
Table 2 Comparison of results for batch reactor consecutive reaction problem S. no. No. of control stages (D) Jbest Using TDE 1 2 3 4 5 6 10 20 25 40 50 80 0.610070 0.610453 0.610535 0.610663 0.610707 0.610775
k1 k2

stages faster than DE. However, percentage of runs converged to global optima in case of DE and TDE are 100% for all control stages. Also, the control prole obtained using DE and TDE is same and matches with that reported in Dadebo and Mcauley (1995). 3.1.3. Plug ow reactor catalyst blend problem Optimization of plug ow reactor studied by Gunn and Thomas (1965) involving the blending of catalyst has been analyzed. This problem was solved analytically by Jackson (1968) and numerically by Logsdon and Biegler (1989). The model describing the system dynamics of the reactant and product concentrations is given in Dadebo and Mcauley (1995). The control variable is discretized into D number of intervals or control stages as reported in (Dadebo & Mcauley, 1995). The key parameters used are F = 0.5, CR = 0.8, NP = 2D and Mt = 0.32. The results of 50 different runs of each algorithm for 40 control stages are plotted in Fig. 3. The overall accuracy of 50-f values is quantied by their low average of 0.00024938 in the range 0.0002480.00025 for DE and 0.00025584 in the range 0.0002520.000261 for TDE. The precision of results is quantied by standard deviation of 5.30306 1007 for DE and 2.17931 1006 for TDE. Table 3 shows the comparison of results for both the algorithms using different number of control stages. CPU-time is the average of 50 different runs. From Table 3, it is clear that for all control stages the CPU-time required to obtain global optimum using TDE is signicantly less than that of obtained using

Average CPU-time (s) Using DE 0.610070 0.610453 0.610536 0.610664 0.610708 0.610775 From literature 0.61007 0.610453 0.610535 0.610664 0.610708 0.610775 TDE 0.5125 1.44186 2.25062 3.1197 4.71592 8.13156 DE 0.812 2.70342 5.55876 11.93374 28.83156 73.39197826

R. Angira, A. Santosh / Computers and Chemical Engineering 31 (2007) 10551063 Table 3 Comparison of results for plug ow reactor catalyst blend problem S. no. No. of control stages (D) Jbest Using TDE 1 2 20 40 0.475269 0.476826 Using DE 0.475269 0.476827 From literature 0.475272 0.476946 Average CPU-time (s) TDE 0.26092 1.33562 DE 0.63282 3.52624

1059

DE algorithm. TDE is approximately 2.42 times for 20 control stages and 2.64 times for 40 control stages faster than DE. However, percentage of runs converged to global optimum in cases of DE and TDE are 100 and 95%, respectively, for all control stages. The control prole obtained using DE is nearly same as that obtained using TDE. The three problems discussed above have also been solved by Logsdon and Biegler (1989), Dadebo and Mcauley (1995), and Rajesh et al. (2001). The results from literature are shown in Table 4. A direct comparison of CPU-time cannot be done, as the computational platforms are different. However, a meaningful comparison can be made by considering a factor based on the ratio of clock speed of computers used. A factor of 100 (high enough) is considered for comparing the CPU-time with that reported in Dadebo and Mcauley (1995). If the Problems 3.1.1, 3.1.2, and 3.1.3 would have been solved on 486/33 Pakard Bell system (33 MHz) using TDE algorithm, they might have taken a CPU-time of 33.7, 51.25, and 26.09 s, respectively. This is much less than that of 126, 105, and 595.6 s, respectively, for Problems 3.1.1, 3.1.2, and 3.1.3 using IDP. It is clear from Table 4 that the objective function value obtained using DE and TDE is same as that obtained by Dadebo and Mcauley (1995), and Logsdon and Biegler (1989) for all the three problems discussed above. Also, it is to be noted that a better objective function value is obtained using DE and TDE than that obtained by Rajesh et al. (2001) using ant algorithm assuming piecewise linear control prole with a grid size of ve. Since the speed of the system is not reported by Rajesh et al. (2001), a meaningful comparison of CPU-time cannot be done. 3.1.4. Bifunctional catalyst problem (hydroisomerization of methylcyclopentane) This problem is multimodal in nature for which more than 200 solutions are reported in literature (Floudas et al., 1999). Also, it has been studied by Luus, Dittrich, & Keil (1992), Bojkov and
Table 4 Results from literature Problem no. 3.1.1. Method

Fig. 4. Fractional difference variation for bifunctional catalyst problem (10 control stages).

Luus (1993), and Luus and Bojkov (1994). The model equations describing the system dynamics are given in literature (Floudas et al., 1999; Upreti, 2004). The control function is discretized into number of intervals or control stages (D = 10) as reported in (Upreti, 2004). The key parameters of DE and TDE are taken as, F = 0.99, CR = 0.85 and NP = 30 and Mt = 0.10. The results of 50 different runs of both the algorithms are plotted in Fig. 4. The best objective function value is reported in literature [Jbest = 1.00942 1002 ; Luus et al. (1992)]. The overall accuracy of 50-f values is quantied by their low average of 0.00085614 in the range 0.000060.005524 for DE, and 0.0015949 in the range 0.0000560.014937 for TDE. The precision of results is quantied by standard deviation of 0.001893258 for DE, and 0.00296794 for TDE. Table 5 shows the comparison of results for both the algorithms. CPU-time is the average of 50 different runs. From Table 5, it is clear that DE and TDE are able to obtain global optimum value but TDE took 28.91% less CPU-time as compared to DE. However, number of runs converged to global optima in

Objective function value With in 0.2% of optimum 0.57353 0.57284 0.610070 0.610767 0.61045 0.475272 0.476946 0.47615

Average CPU-time (s) 126 (486/33 Packard Bell system) Not reported 0.07 (Sun Enterprise 450 server) 105 (486/33 Packard Bell system) Not reported 0.07 (Sun Enterprise 450 server) 595.6 (486/33 Packard Bell system) Not reported 0.24 (Sun Enterprise 450 server)

Iterative dynamic programming (Dadebo & Mcauley, 1995) Modied collocation based NLP formulation (Logsdon & Biegler, 1989) Ant colony algorithm (Rajesh et al., 2001) Iterative dynamic programming (Dadebo & Mcauley, 1995) Modied collocation based NLP formulation (Logsdon & Biegler, 1989) Ant colony algorithm (Rajesh et al., 2001) Iterative dynamic programming (Dadebo & Mcauley, 1995) Modied collocation based NLP formulation (Logsdon & Biegler, 1989) Ant colony algorithm (Rajesh et al., 2001)

3.1.2.

3.1.3.

1060

R. Angira, A. Santosh / Computers and Chemical Engineering 31 (2007) 10551063

Table 5 Comparison of results for bifunctional catalyst problem Method DE (present study) TDE (present study) Variant of GA (Upreti, 2004) Sequential quadratic programming (Luus et al., 1992) Iterative dynamic programming (Luus et al., 1992) Iterative dynamic programming (Luus & Bojkov, 1994) Starting point Random Random Random Random u = 0.75 for all stages u = 0.75 for all stages Jbest 0.0100940 0.0100940 0.0100937 0.0100527 0.0100942 0.0100942 Average CPU-time (s) 247.52156 (on IBM computer) 175.9662 (on IBM computer) 166 (Pentium III/750 MHz with 192 MB RAM) 185 (for 100 starting points on CRAY XMP/24 digital computer) 360960 (On CRAY XMP/24 digital computer) 6 (Pentium II/350 MHz)

case of DE and TDE are 86 and 76%, respectively. The control proles for 10 stages using DE and TDE algorithms are found to be similar. It can also be seen from Table 5 that DE and TDE are able to locate a better objective function value as compared to the genetic algorithm of Upreti (2004) but require more computational efforts. Iterative dynamic programming of Luus and Bojkov (1994) took least computational time to locate the global optimum for this problem. 3.1.5. Fed batch bioreactor (production of secreted protein) This problem deals with the optimal production of secreted protein in a fed batch reactor. It is originally formulated by Park and Ramirez (1988) and it has also been considered by other researchers (Luus & Hennessy, 1999; Upreti, 2004). The model equations describing the system dynamics are given in literature (Park & Ramirez, 1988; Upreti, 2004). The control function is discretized into number of intervals or control stages (D = 45) as in (Upreti, 2004). The key parameters of DE and TDE are taken as, F = 0.99, CR = 0.85, and NP = 30 and Mt = 0.21. The results of 50 different runs of both the algorithms are shown in Fig. 5. The best objective function value reported in literature (Luus & Hennessy, 1999) is Jbest = 32.686867. The overall accuracy of 50-f values is quantied by their low average of 0.00014256 in the range 0.0000730.003541 for DE, and 0.00012268 in the range 0.0000770.000563 for TDE. The precision of results is quantied by standard deviation of 0.000490421 for DE, and 9.43263 1005 for TDE. Table 6 shows the comparison of results for both the algorithms. CPU-time is the average of 50 different runs. From Table 6, it is clear that both DE and TDE are able to obtain global optimum but TDE took 25.30% less CPU-time as compared to DE. However, percentage of runs converged to global optima in case of DE and TDE are 98 and 64%, respectively. The control proles obtained using both the algorithms are slightly different (Fig. 6).
Table 6 Comparison of results for fed batch bioreactor Method DE (present study) TDE (present study) Variant of GA (Upreti, 2004) Iterative dynamic programming (Luus & Hennessy, 1999) Direct search (Luus & Hennessy, 1999) Starting point Random Random Random u = 1 for all stages Jbest 32.684494 32.684338 32.59277 32.686867 32.686867 Average CPU-time (s) 308.1328 (on IBM computer) 230.1591 (on IBM computer) 1244 (Pentium III/750 MHz with 192 MB RAM) 35847391 (on Pentium/120) less than 36968 (on Pentium/120)

Fig. 5. Fractional difference variation for fed batch bioreactor (45 control stages).

Fig. 6. Optimal control prole for fed batch bioreactor (45 control stages).

It is evident from Table 6 that DE and TDE algorithms perform better than the genetic algorithm of Upreti (2004), both in terms of computational cost and ability to locate global optimum. A meaningful comparison can be done by considering a factor of 4 (high enough) based on the ratio of the clock speed

R. Angira, A. Santosh / Computers and Chemical Engineering 31 (2007) 10551063

1061

of computers used, i.e. if the same problem would have been solved on P-III/750 MHz using TDE, it might have taken a CPUtime of 921 s as compared to 1244 s using genetic algorithm of Upreti (2004). DE and TDE algorithms seem to be faster than the direct search and iterative dynamic programming of Luus and Hennessy (1999). However, direct search and iterative dynamic programming techniques are able to locate a slightly better objective function value as compared to other methods shown in Table 6. 3.2. Parameter estimation problems For the numerical solution of the problems considered in the present study, the state parameters are discretized in to r stages of known experimental state parameter data. The dynamic system is then integrated using RungeKutta fourth-order method, in each r stage to evaluate the objective function and the constraints. 3.2.1. A rst-order irreversible liquid-phase reaction This example is a parameter estimation problem with two parameters and two differential equations in the constraints. It has been solved by Tjoa and Biegler (1991), Esposito and Floudas (2000), Papamichail and Adjiman (2004). Also it appears in Floudas et al. (1999). The reader is referred to Floudas et al. (1999) for details of the problem formulation. The key parameters of DE are taken as, F = 0.5, CR = 0.8, and NP = 2D. A Mt value of 0.9 is taken for TDE. The results of 50 different runs using both the algorithms are plotted in Fig. 7. The best objective function value reported in Floudas et al. (1999) is 1.18584 1006 . The overall accuracy of 50-f values is quantied by their low average of 0.000004 for both the algorithms. The precision of results is quantied by standard deviation of zero for both DE and TDE algorithms. Table 7 shows the comparison of results for both the algorithms. From Table 7, it is clear that both DE and TDE algorithms are converging to global optimum value. The CPU-time required to obtain global optimum using TDE algorithm is nearly same as required by DE. Fig. 8 shows the experimental data points and the obtained data points using DE and TDE algorithms. From Fig. 8, it is clear that obtained state variable data is matching with that experimen-

Table 7 Comparison of results for a rst-order irreversible liquid-phase reaction Method DE (present study) TDE (present study) Spatial branch and bound (Papamichail & Adjiman, 2004) Jbest 1.185845 l006 1.18562 1006 1.185845 l006 Average CPU-time (s) 0.47218 (on IBM computer) 0.45626 (on IBM computer) 767a

a Ultra SPARC-II CPU/2 360 MHz, constant afne bounds for convex relaxation of equality constraint.

tal data as reported in Floudas et al. (1999) and Papamichail and Adjiman (2004). The global optimum parameters obtained by DE and TDE algorithms in all the fty different runs are same, i.e. k1 = 5.003487 and k2 = 1.00. 3.2.2. Catalytic cracking of gas oil This is also a parameter estimation problem with three parameters and two differential equations in the constraints. It has been solved by Tjoa and Biegler (1991), Esposito and Floudas (2000), Papamichail and Adjiman (2004). The problem formulation is same as that reported in literature (Esposito & Floudas, 2000; Papamichail & Adjiman, 2004).

Fig. 7. Fractional difference variation for a rst-order irreversible liquid-phase reaction.

Fig. 8. Experimental points and state variable trajectories for a rst-order irreversible liquid-phase reaction.

1062

R. Angira, A. Santosh / Computers and Chemical Engineering 31 (2007) 10551063

Fig. 9. Fractional difference variation for catalytic cracking of gas oil problem.

The key parameters of DE are taken as, F = 0.5, CR = 0.8, NP = 2D and Mt = 0.01 for TDE. The results of 50 different runs of both the algorithms are plotted in Fig. 9. The best objective function value reported in Floudas et al. (1999) is 2.65567 1003 . The overall accuracy of 50-f values is quantied by their low average of 0.000001 for both algorithms. The precision of results is quantied by standard deviation of zero for both algorithms. Table 8 shows comparison of results of all three algorithms. From Table 8, it is clear that both algorithms are converging to global optimum value. The CPU-time required to obtain global optimum using TDE algorithm is more than that required by DE. Fig. 10 shows the experimental data points, results from Floudas et al. (1999) and the data points obtained in the present study. From Fig. 10, it is clear that obtained state variable data is matching with that literature data points (Floudas et al., 1999). The global optimum parameters obtained by DE and TDE algorithms in all the 50 different runs are same, i.e. k1 = 12.21401, k2 = 7.979833, and k3 = 2.221622. For the above two parameter estimation problems, the CPUtime taken by DE and TDE algorithms is much less than that of spatial branch and bound global optimization algorithm (Tables 7 and 8). Of course the CPU-times cannot be compared directly because of different computational platforms (PentiumIV (2.4 GHz) using Microsoft Visual C++ 6.0, compiler in the present study and Ultra SPARC-II CPU (2 360 MHz, 512
Table 8 Comparison of results for catalytic cracking of gas oil Method DE (present study) TDE (present study) Spatial branch and bound (Papamichail & Adjiman, 2004) Jbest 2.655666 1003 2.65566604 l 1003 2.65567 1003 Average CPU-time (s) 1.16592 (on IBM computer) 1.8078 (on IBM computer) 16729a

Fig. 10. Experimental points and state variable trajectories of catalytic cracking of gas oil.

MB RAM) using MATLAB 5.3 by Papamichail and Adjiman (2004)). However, a meaningful comparison can be done after considering a factor of 10 (high enough) based on the ratio of the clock speed of computers used, i.e. if the Problems 3.2.1 and 3.2.2 would have been solved on Ultra SPARC-II CPU (2 360 MHz, 512 MB RAM) using DE and TDE, they might have taken 10 times more of CPU-time than at Pentium-IV (2.4 GHz). Even then the CPU-time taken by DE and TDE is signicantly less than that of spatial branch and bound global optimization algorithm. This clearly indicates the better performance of DE and TDE as compared to that of spatial branch and bound global optimization algorithm for the parameter estimation problems considered in the present study. 4. Conclusions In this paper TDE algorithm is evaluated for solving ve nonlinear optimal control problems and two parameter estimation problems from chemical engineering. Both DE and TDE algorithms are able to obtain global optimum with nearly 100% convergence overall the 50 different executions of the algorithms. The performance of TDE algorithm is compared with that of DE and other methods from literature. The TDE algorithm is found to be efcient and signicantly faster than the DE, particularly for the large number of control stages. Also, it is found to be computationally efcient than the spatial branch and bound method for the problem considered in the present

a Ultra SPARC-II CPU/2 360 MHz, constant bounds for convex relaxation of equality constraint.

R. Angira, A. Santosh / Computers and Chemical Engineering 31 (2007) 10551063

1063

study. TDE is found to locate a better objective function value than that using genetic algorithm of Upreti for the two optimal control problems considered in this study. Hence, TDE seems to be a promising algorithm to solve dynamic optimization problems where computation of objective function is quite expensive. References
Angira, R. (2005). Evolutionary computation for optimization of selected nonlinear chemical processes. Ph. D. Thesis, Birla Institute of Technology & Science (BITS), Pilani, India. Angira, R., & Alladwar, S. (2005). Optimal control of chemical engineering processes using trigonometric differential evolution algorithm. In Proceedings of International Symposium & 58th Annual Session of IIChE in association with International Partners (CHEMCON-2005). Angira, R., & Babu, B. V. (2003). Evolutionary computation for global optimization of non-linear chemical engineering processes. In Proceedings of International Symposium on Process Systems Engineering and Control (ISPSEC 03)-For Productivity Enhancement through Design and Optimization. Angira, R., & Babu, B. V. (2005). Optimization of non-linear chemical processes using modied differential evolution (MDE). In Proceedings of the 2nd Indian International Conference on Articial Intelligence (IICAI-2005) (pp. 911923,). Angira, R., & Babu, B. V. (2006). Optimization of process synthesis and design problems: A modied differential evolution approach. Chemical Engineering Science, 61, 47074721. Babu, B. V., & Angira, R. (2006). Modied differential evolution (MDE) for optimization of nonlinear chemical processes. Computers & Chemical Engineering, 30, 9891002. Barton, P. I., Allgor, R. J., Feehery, W. F., & Galan, S. (1998). Dynamic optimization in a discontinuous world. Industrial & Engineering Chemistry Research, 37, 966981. Bojkov, B., & Luus, R. (1993). Evaluation of the parameters used in iterative dynamic programming. Canadian Journal of Chemical Engineering, 780. Carrasco, E. F., & Banga, J. R. (1997). Dynamic optimization of Batch reactors using adaptive stochastic algorithms. Industrial & Engineering Chemistry Research, 36, 2252. Chiou, J. P., & Wang, F. S. (1999). Hybrid method of evolutionary algorithms for static and dynamic optimization problems with application to fed batch fermentation process. Computers & Chemical Engineering, 23(9), 12771291. Corne, D., Dorigo, M., & Glover, F. (1999). New ideas in optimization. England, UK: McGraw-Hill, Berkshire. Dadebo, S. A., & Mcauley, K. B. (1995). Dynamic optimization of constrained chemical engineering problems using dynamic programming. Computers & Chemical Engineering, 19, 513525. Esposito, W. R., & Floudas, C. A. (2000). Global optimization for the parameter estimation of differential algebraic systems. Industrial & Engineering Chemistry Research, 39, 12911310. Fan, H. Y., & Lampinen, J. (2003). A trigonometric mutation operation to differential evolution. Journal of Global Optimization, 27, 105129. Floudas, C. A. (2000). Deterministic global optimization: Theory, methods and applications. In Series in nonconvex optimization and its applications. Dordrecht: Kluwer Academic Publishers. Floudas, C. A., Pardalos, P. M., Adjiman, C. S., Esposito, W. R., Gumiis, Z. H., Harding, S. T., Klepeis, J. L., Meyer, C. A., & Schweiger, C. A. (1999). Handbook of test problems in local and global optimization. In Series in nonconvex optimization and its applications. Dordrecht: Kluwer Academic Publishers. Gunn, D. J., & Thomas, W. J. (1965). Mass transport and chemical reaction in multifunctional catalyst systems. Chemical Engineering Science, 20, 89100.

Jackson, R. (1968). Optimal use of mixed catalysts for two successive chemical reactions. Journal of Optimization Theory & Application, 2(1). Lim, H. C., Tayeb, Y. J., Modak, J. M., & Bonte, P. (1986). Computational algorithms for optimal feed rates for a class of fed-batch fermentation: Numerical results for penicillin and cell mass production. Biotechnology and Bioengineering, 28, 1408. Logsdon, J. S., & Biegler, L. T. (1989). Accurate solution of differential algebraic equations. Industrial & Engineering Chemistry Research, 2, 1628. Lucia, A., Bellows, M. L., & Octavio, L. M. (2005). Global optimization of ordinary differential equations models. In L. Puijanier & A. Espuna (Eds.), ESCAPE-15, Vol. 20A (p. 115). Lucia, A., DiMaggio, P. A., & Depa, P. (2004). A geometric terrain methodology for global optimization. Journal of Global Optimization, 29, 297. Luus, R. (1994). Optimal control of batch reactors by iterative dynamic programming. Journal of Process control, 4(4), 218. Luus, R., & Bojkov, B. (1994). Global optimization of the bifunctional catalyst problem. Canadian Journal of Chemical Engineering, 72, 160. Luus, R., & Hennessy, D. (1999). Optimization of fed batch reactors by the Luss-Jaakola optimization procedure. Industrial & Engineering Chemistry Research, 35(5), 1948. Luus, R., & Jaakola, T. H. I. (1973). Optimization by direct search and systematic reduction of the size of search region. AIChE Journal, 19, 760766. Luus, R., Dittrich, J., & Keil, F. (1992). Multiplicity of solutions in the optimization of a bifunctional catalyst blend in a tubular reactor. Canadian Journal of Chemical Engineering, 70, 780. Onwubolu, G. C., & Babu, B. V. (2004). New optimization techniques in engineering. Heidelberg, Germany: Springer. Papamichail, I., & Adjiman, C. S. (2004). Global optimization of dynamic systems. Computers & Chemical Engineering, 28, 403415. Park, S., & Ramirez, W. F. (1988). Optimal production of secreted protein in fed batch reactors. AIChE Journal, 34(8), 15501558. Pham, Q. T. (1998). Dynamic optimization of chemical processes by an evolutionary method. Computers & Chemical Engineering, 22, 1089 1097. Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. (2002). Numerical recipes in C++. In The art of scientic computing (2nd ed.). New York: Cambridge University Press, pp. 719727. Rajesh, J., Gupta, K., Kusumakar, H., Jayaraman, V. K., & Kulkarni, B. D. (2001). Dynamic optimization of chemical processes using ant colony framework. Computers and Chemistry, 25, 583595. Sarkar, D., & Modak, J. M. (2003). Optimization of fed batch bioreactors using genetic algorithms. Chemical Engineering Science, 58, 22832296. Storn, R., & Price, K. (1995). DE-A Simple and Efcient Adaptive Scheme for Global Optimization over Continuous Space, Technical Report TR95-012, ICSI, March, Available via the Internet: ftp.icsi.berkeley.edu/pub/ techreports/1995/tr-95-012.ps.Z. Storn, R., & Price, K. (1996). Minimizing the real function of the ICEC96 contest by DE. In IEEE International Conference on Evolutionary Computation (pp. 842844). Tang, W., Zhang, L., Linninger, A. A., Tranter, R. S., & Brezinsky, K. (2005). Solving kinetic inversion problems via a physically bounded GaussNewton (PGN) method. Industrial & Engineering Chemistry Research, 44, 3626. Tjoa, I. B., & Biegler, L. T. (1991). Simultaneous solution and optimization strategies for parameter estimation of differential-algebraic equation systems. Industrial & Engineering Chemistry Research, 30, 376385. Upreti, S. R. (2004). A new robust technique for optimal control of chemical engineering processes. Computers & Chemical Engineering, 28, 1325 1336. Wang, F. S., & Chiou, J. P. (1997). Optimal control and optimal time location problems of differential algebraic systems by differential evolution. Industrial & Engineering Chemistry Research, 36, 53485357.