Mimo system-order-reduction-using-real-coded-genetic-algorithm

Mimo system-order-reduction-using-real-coded-genetic-algorithm

• 1.
  World Academy ofScience, Engineering and Technology International Journal of Electrical, Electronic Science and Engineering Vol:5 No:3, 2011 MIMO System Order Reduction Using RealCoded Genetic Algorithm International Science Index 51, 2011 waset.org/publications/4570 Swadhin Ku. Mishra, Sidhartha Panda, Simanchala Padhy and C. Ardil Abstract—In this paper, real-coded genetic algorithm (RCGA) optimization technique has been applied for large-scale linear dynamic multi-input-multi-output (MIMO) system. The method is based on error minimization technique where the integral square error between the transient responses of original and reduced order models has been minimized by RCGA. The reduction procedure is simple computer oriented and the approach is comparable in quality with the other well-known reduction techniques. Also, the proposed method guarantees stability of the reduced model if the original high-order MIMO system is stable. The proposed approach of MIMO system order reduction is illustrated with the help of an example and the results are compared with the recently published other well-known reduction techniques to show its superiority. Keywords—Multi-input-multi-output (MIMO) system. Model order reduction. Integral squared error (ISE). Real-coded genetic algorithm. I. INTRODUCTION I N system theory the approximation of higher order system by lower order models is one of the important challenges. Because by using the reduced lower order model, the implementation of analysis, simulation and various system designs become easier. Various order-reduction methods are available for linear continuous time domain system as well as systems in frequency domain. Further, the extension of singleinput single-output (SISO) methods to reduce multi-input multi-output (MIMO) systems has also been carried out in [1][3]. Various order reduction methods are available in [4]-[8] based on the minimization of the integral square error (ISE) criterion. However, a familiar aspect in the methods explained in [4]-[7] is that the denominator coefficients values of the low order system (LOS) are selected arbitrarily by some stability preserving methods such as dominant pole, Routh Swadhin Kumar Mishra is working as Assistant Professor in Electronics and Communication Engg. Department, NIST Berhampur, Orissa, India, Pin: 761008 (e-mail: swadhin.mishra@gmail.com ). S. Panda is working as a Professor in the Department of Electrical and Electronics Engineering, NIST, Berhampur, Orissa, India, Pin: 761008. (email: panda_sidhartha@rediffmail.com ). Simanchala Padhy is working as Engineering Assistant, at HPT, Berhampur, Prasar Bharati Broadcasting Corporation of India (e-mail : simanchala.ddair@gmail.com) C. Ardil is with National Academy of Aviation, AZ1045, Baku, Azerbaijan, Bina, 25th km, NAA (e-mail: cemalardil@gmail.com). approximation methods, etc. and then the numerator coefficients of the LOS are determined by minimization of the ISE. In [8], Howitt and Luss recommended a procedure, in which both the numerator and denominator coefficients are considered to be free parameters and are chosen to minimize the ISE in impulse or step responses. These days, Genetic algorithm (GA) is becoming popular to solve the optimization problems in different fields of application mainly because of their robustness in finding an optimal solution and ability to provide a near optimal solution close to a global minimum. Unlike strict mathematical methods, the GA does not require the condition that the variables in the optimization problem be continuous and different; it only requires that the problem to be solved can be computed. The present attempt is towards evolving a new algorithm for order reduction, where all the numerator and denominator parameters are considered to be free parameters and the error minimization by RCGA is employed to optimize these parameters. The proposed algorithm consists of searching all the parameters by minimizing the integral square error between the transient responses of original and LOS using RCGA. The proposed approach is illustrated with the help of an example and the results are compared with the recently published techniques. II. REAL-CODED GENERIC ALGORITHM Genetic algorithm (GA) has been used to solve difficult engineering problems that are complex and difficult to solve by conventional optimization methods. GA maintains and manipulates a population of solutions and implements a survival of the fittest strategy in their search for better solutions. The fittest individuals of any population tend to reproduce and survive to the next generation thus improving successive generations. The inferior individuals can also survive and reproduce. Implementation of GA requires the determination of six fundamental issues: chromosome representation, selection function, the genetic operators, initialization, termination and evaluation function. Brief descriptions about these issues are provided in the following sections. A. Chromosome representation Chromosome representation scheme determines how the problem is structured in the GA and also determines the genetic operators that are used. Each individual or chromosome is made up of a sequence of genes. Various types of representations of an individual or chromosome are: binary 17
• 2.
  World Academy ofScience, Engineering and Technology International Journal of Electrical, Electronic Science and Engineering Vol:5 No:3, 2011 International Science Index 51, 2011 waset.org/publications/4570 digits, floating point numbers, integers, real values, matrices, etc. Generally natural representations are more efficient and produce better solutions. Real-coded representation is more efficient in terms of CPU time and offers higher precision with more consistent results. B. Selection Function To produce successive generations, selection of individuals plays a very significant role in a genetic algorithm. The selection function determines which of the individuals will survive and move on to the next generation. A probabilistic selection is performed based upon the individual’s fitness such that the superior individuals have more chance of being selected. There are several schemes for the selection process: roulette wheel selection and its extensions, scaling techniques, tournament, normal geometric, elitist models and ranking methods. The selection approach assigns a probability of selection Pj to each individuals based on its fitness value. In the present study, normalized geometric selection function has been used. In normalized geometric ranking, the probability of selecting an individual Pi is defined as: Pi = q ' (1 − q )r −1 (1) q' = q 1 − (1 − q ) p (2) where, q = probability of selecting the best individual r = rank of the individual (with best equals 1) P = population size C. Genetic Operator The basic search mechanism of the GA is provided by the genetic operators. There are two basic types of operators: crossover and mutation. These operators are used to produce new solutions based on existing solutions in the population. Crossover takes two individuals to be parents and produces two new individuals while mutation alters one individual to produce a single new solution. The following genetic operators are usually employed: simple crossover, arithmetic crossover and heuristic crossover as crossover operator and uniform mutation, non-uniform mutation, multi-non-uniform mutation, boundary mutation as mutation operator. Arithmetic crossover and non-uniform mutation are employed in the present study as genetic operators. Crossover generates a random number r from a uniform distribution from 1 to m and creates two new individuals by using equations: ⎧ xi , xi' = ⎨ ⎩ yi , ⎧ yi , yi' = ⎨ ⎩ xi , if i < r (3) otherwise if i < r ′ X = r X + (1 − r )Y ′ Y = rY + (1 − r ) X (5) Non-uniform mutation randomly selects one variable j and sets it equal to an non-uniform random number xi' = xi + (bi − xi ) f (G ) if r1 < 0.5 xi' = xi + ( xi + ai ) f (G ) if r1 ≥ 0.5 xi (6) otherwise where ⎛ ⎛ G ⎞⎞ ⎟⎟ f (G ) = ⎜ r2 ⎜1 − ⎟⎟ ⎜ ⎜ G max ⎠ ⎠ ⎝ ⎝ b (7) r1, r2 = uniform random numbers between 0 and 1. G = current generation. Gmax = maximum number of generations. b = shape parameter D. Initialization, evaluation function and stopping criteria An initial population is needed to start the genetic algorithm procedure. The initial population can be randomly generated or can be taken from other methods. Evaluation functions or objective functions of many forms can be used in a GA so that the function can map the population into a partially ordered set. The GA moves from generation to generation until a stopping criterion is met. The stopping criterion could be maximum number of generations, population convergence criteria, lack of improvement in the best solution over a specified number of generations or target value for the objective function. III. DESCRIPTION OF THE PROPOSED ALGORITHM. Let the transfer function matrix of the HOS of order 'n' having ‘p’ inputs and ‘m’ outputs are: ⎡ a11 ( s ) a12 ( s ) ⎢ 1 ⎢ a21 ( s ) a22 ( s ) [G ( s )] = M Dn ( s ) ⎢ M ⎢ ⎢am1 ( s ) am 2 ( s ) ⎣ L a1 p ( s ) ⎤ L a2 p ( s ) ⎥ ⎥ (8) L M ⎥ ⎥ L amp ( s )⎥ ⎦ or, [G(s)] = [gij(s)] i=1, 2… m and j=1, 2… p is a m x p matrix. The general form of gij(s) is taken as: aij ( s ) a0 + a1s + a2 s 2 + .... + an −1s n −1 gij ( s ) = = Dn ( s ) b0 + b1s + b2 s 2 + .... + bn −1s n −1 + s n (4) (9) otherwise Arithmetic crossover produces two complimentary linear combinations of the parents, where r = U (0, 1): D(s) is the nth order denominator polynomial of the transmission matrix for the HOS. 18
• 3.
  World Academy ofScience, Engineering and Technology International Journal of Electrical, Electronic Science and Engineering Vol:5 No:3, 2011 ⎡ 2( s + 5) ⎢ 1 [G( s)] = ⎢ ( s +s )(s +)10) ( + 10 ⎢ ⎢ ( s + 1)( s + 20) ⎣ Let, the transfer function matrix of the HOS of order 'n' having ‘p’ inputs and ‘m’ outputs to be synthesized is: ⎡ b11( s ) b12 ( s ) ⎢ 1 ⎢ b21( s ) b22 ( s ) [ R( s )] = M Dr ( s ) ⎢ M ⎢ ⎢bm1 ( s ) bm 2 ( s ) ⎣ L b1 p ( s ) ⎤ L b2 p ( s ) ⎥ ⎥ (10) L M ⎥ ⎥ L bmp ( s )⎥ ⎦ International Science Index 51, 2011 waset.org/publications/4570 or, [R(s)] = [rij(s)] i=1, 2… m and j=1, 2… p is a m x p matrix. The general form of rij(s) is taken as: bij ( s) μ + μ1s + μ2 s 2 + .... + μ r −1s r −1 rij ( s) = = 0 Dr ( s) b0 + b1s + b2 s 2 + .... + br −1s r −1 + s r ( s + 4) ⎤ 1 ⎡ a11 ( s) a12 ( s ) ⎤ ( s + 2)( s + 5) ⎥ ⎥= ( s + 6) ⎥ D6 ( s) ⎢a21( s ) a22 ( s ) ⎥ ⎣ ⎦ ( s + 2)( s + 3) ⎥ ⎦ (13) where D6(s) is the common denominator polynomial. As can be verified the MIMO system is a 6th order system and is given as D6 ( s ) = ( s + 1)( s + 2)( s + 3)( s + 5)( s + 10)( s + 20) = 6000 + 13100 s + 10060 s 2 + 3491s 3 + 571s 4 + 41s 5 + s 6 and a11 ( s ) = 6000 + 7700 s + 3610 s 2 + 762 s 3 + 70 s 4 + 2 s 5 (11) Dr(s) is the rth order denominator polynomial of the transmission matrix for the reduced LOS. The reduction process from an nth order HOS to a reduced rth order LOS consists of the following two steps : a12 ( s ) = 2400 + 4100s + 2182s 2 + 459 s 3 + 38s 4 + s 5 a21 ( s ) = 3000 + 3700 s + 1650 s 2 + 331s 3 + 30 s 4 + s 5 a22 ( s) = 6000 + 9100 s + 3660 s 2 + 601s 3 + 41s 4 + s 5 Step-1 The coefficients of the denominator of the reduced LOS are found out by employing GA. The integral square error (ISE) is found out between the transient responses between the HOS and LOS and is given as 2 ∞ E = ∫ [ y (t ) − yr (t )] dt The proposed algorithm is applied to minimize the function E (ISE). The ISE is calculated for each element (rij(s)) of the transfer function matrix of the LOS and it is given by Eqn. (14) ∞ [ where i=1,2 ; j=1,2; and gij(t) and rij(t) are the unit step response of the original and reduced order model, respectively. The general form of the transfer function matrix of the reduced second order LOS is taken as [R( s)] = Step-2 IV. NUMERIC EXAMPLE To demonstrate the proposed method, one numeric example of a higher order system (6th order) is taken from literature[9][10] and the proposed algorithm is employed to obtain a second-order reduced model. Consider a system having two inputs and two outputs with transfer matrix of (14) 0 0 Once the denominator coefficients of the reduced LOS are found out, GA is used to find out the numerator coefficients of LOS. Here also the Real Coded Generic Algorithm is implemented in such a way to minimize the ISE by keeping the values of the denominator coefficients, which are found out in step 1, of the reduced LOS constant. ] E = ∫ gij (t ) − rij (t ) dt (12) where y(t) and yr(t) are the unit step responses of original G(s) and reduced R(s) order systems. Real Coded Generic Algorithm is employed to minimize the integral square error (ISE) in order to find out the denominator coefficients of the LOS. 2 1 ⎡ b11 ( s ) b12 ( s ) ⎤ Dr ( s ) ⎢b21 ( s ) b22 ( s )⎥ ⎣ ⎦ (15) Where Dr(s) is the common denominator polynomial of the reduced model. For the implementation of RCGA normal geometric selection is employed which is a ranking selection function based on the normalized geometric distribution. Arithmetic crossover takes two parents and performs an interpolation along the line formed by the two parents. Non uniform mutation changes one of the parameters of the parent based on a non-uniform probability distribution. This Gaussian distribution starts wide, and narrows to a point distribution as the current generation approaches the maximum generation. The follow chart of proposed approach is shown in Fig. 1. The reduced 2nd order common denominator polynomial is found by implementing our method as: 19
• 4.
  World Academy ofScience, Engineering and Technology International Journal of Electrical, Electronic Science and Engineering Vol:5 No:3, 2011 Dr ( s ) = 1.0459s 2 + 4.1138s + 2.9920 0.5 and 0.4 Amplitude b11( s ) = 1.3935s + 2.9918 b12 ( s ) = 1.0773s + 1.1985 b21( s) = 0.6247 s + 1.4963 b22 ( s ) = 1.8935s + 3.0069 0.2 0.1 The step responses of original and reduced order models are compared in Figs. 2(a-d). A comparison of proposed method with the other well-known order-reduction techniques available in literature is given in the Table 1. The comparison is made by calculating the ISE given by Eqn. (14) pertaining to a step input. The ISE is calculated for each element of the transfer function matrix of the LOS with respect to the corresponding original system. 0 Original 6th order system Reduced 2nd order system 0 2 4 Time (sec) 6 8 (b) 0.6 Amplitude 0.5 Start 0.4 0.3 0.2 Original 6th order system Reduced 2nd order system 0.1 Specify the parameters for GA 0 0 2 Generate initial population 4 Time (sec) 6 8 (c) Gen.=1 1 Time-domain simulation Amplitude 0.8 Find the fittness of each individual in the current population 0.6 0.4 Gen.=Gen.+1 Gen. > Max. Gen.? 0.2 Stop Yes 0 No 0 2 4 Time (sec) 6 8 Fig. 2 Comparison of step response (a) Output 1 for input 1 (b) Output 1 for input 2 (c) Output 2 for input 1 (d) Output 2 for input 2 Fig. 1 Flowchart of the proposed approach V. CONCLUSION 1 0.8 0.6 0.4 0.2 0 Original 6th order system Reduced 2nd order system (d) Apply GA operators: selection,crossover and mutation Amplitude International Science Index 51, 2011 waset.org/publications/4570 0.3 Original 6th order system Reduced 2nd order system 0 2 4 Time (sec) (a) 6 8 An approach based on the error minimization technique employing real-coded genetic algorithm has been presented, to derive stable reduced order models for linear time invariant multi-input multi-output dynamic systems. The algorithm has been implemented in MATLAB 7.0.1 on an Inetl Core 2 Duo processor and the computation time is negligible being less than 1 minute. The matching of the step response is assured reasonably well in the method. The ISE in between the transient parts of original and reduced order systems is calculated and compared in the tabular form as given in Tables 1, from which it is clear that the proposed algorithm compares well with the other existing techniques of model order 20
• 5.
  World Academy ofScience, Engineering and Technology International Journal of Electrical, Electronic Science and Engineering Vol:5 No:3, 2011 reduction. The method also preserves the model stability and avoids any steady-state error between the time responses of the original and reduced systems. TABLE I COMPARISON OF REDUCTION METHODS r11 r12 r21 r22 Proposed method Vishwakarma and Prasad [9] Parmar et al. [10] Prasad and Pal [11] Safonov and Chiang [12] Prasad et al. [13] International Science Index 51, 2011 waset.org/publications/4570 Reduction method 4.0656x10-4 7.772x10-5 3.2448x10-5 0.0068 0.001515 7.845x10-5 2.9984x10-4 0.0047 0.014498 0.008744 0.002538 0.015741 0.136484 0.002446 0.040291 0.067902 0.590617 0.037129 0.007328 1.066123 0.030689 0.000256 0.261963 0.021683 REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] S. Mukherjee and R.N. Mishra,“Reduced order modelling of linear multivariable systems using an error minimization technique”, Journalof Franklin Inst., Vol. 325, No. 2 , pp. 235-245,1998. S. S. Lamba, R. Gorez and B. Bandyopadhyay “New reduction technique by step error minimization for multivariable systems”, Int. J. Systems Sci., Vol. 19, No. 6, pp. 999-1009,1988. R. Prasad, A. K. Mittal and S. P. Sharma “ A mixed method for the reduction of multivariable systems”, Journal of Institute of Engineers, India, IE(I) Journal-EL, Vol. 85, pp 177-181,2005. 4. C. Hwang “ Mixed method of Routh and ISE criterion approaches for reduced order modelling of continuous time systems”, Trans. ASME, J. Dyn. Syst. Meas. Control, Vol. 106,pp. 353-356. 1984. S. Mukherjee, and R.N. Mishra, “Order reduction of linear systems using an error minimization technique”, Journalof Franklin Inst., Vol. 323, No. 1, pp. 23-32, 1987. N.N. Puri, and D.P. Lan, “Stable model reduction by impulse response error minimization using Mihailov criterion and Pade’s approximation”, Trans. ASME, J. Dyn. Syst. Meas. Control, Vol. 110, pp. 389-394, 1988. P. Vilbe, and L.C. Calvez, “On order reduction of linear systems using an error minimization technique”, Journal of Franklin Inst., Vol. 327, pp. 513-514, 1990. G.D. Howitt, and R. Luus, “Model reduction by minimization of integral square error performance indices”, Journal of Franklin Inst., Vol. 327, pp. 343-357, 1990. Vishwakarma and Prasad, “MIMO System Reduction Using Modified Pole Clustering and Genetic Algorithm”, Hindawi Pub. Corp. Mod. and Sim. in Engineering, G. Parmar, R. Prasad, and S. Mukherjee, “Order reduction of linear dynamic systems using stability equation method and GA,” International Journal of Computer, Information, and Systems Science, and Engineering, vol. 1, no. 1, pp. 26–32, 2007. R. Prasad and J. Pal, “Use of continued fraction expansion for stable reduction of linearmultivariable systems,” Journal of the Institution of Engineers, vol. 72, pp. 43–47, 1991. M. G. Safonov and R. Y. Chiang, “Model reduction for robust control: a schur relative errormethod,” International Journal of Adaptive Control and Signal Processing, vol. 2, no. 4, pp. 259–272, 1988. R. Prasad, J. Pal, and A. K. Pant, “Multivariable system approximation using polynomial derivatives,” Journal of the Institution of Engineers, vol. 76, pp. 186–188, 1995. Swadhin Kumar Mishra is working as an Asst. Prof. at National institute of Science and Technology (NIST), Berhampur, Orissa. He has done his M.Tech at Indian Institute of Technology, Madras, India in 2009 and his B.E. degree in 2000. Earlier he worked as an Assistance Professor at RIT, Berhampur, as Senior Lecturer at JITM, Paralakhemundi, Orissa, India and as lecturer at SMIT, Berhampur, Orissa, India. His areas of research include MIMO communication and optical communication. Dr. Sidhartha Panda received Ph.D. degree from Indian Institute of Technology (IIT), Roorkee, India in 2008, M.E. degree from VSS University of Technology, (erstwhile UCE, Burla) in 2001 and. B.E. degree from Bangalore University in 1991 all in Electrical Engineering. Presently he is working as a Professor in the EEE department at National Institute of Science and Technology (NIST), Berhampur, Orissa. Earlier he worked as an Associate Professor in KIIT Deemed University and also in various other engineering colleges for about 15 years. He has published about 60 papers in various International Journals. Presently, he is acting as a reviewer of some International Journals namely; IEEE Transactions on Industrial Electronics, Applied Soft Computing (Elsevier), International Journal Electric Power and Energy Systems (Elsevier), International Journal Simulation Modelling Practice and Theory (Elsevier), International Journal of Control and Intelligent Systems (ACTA Press). The biography of Dr Panda has been included “Who's Who in the World”: 2010 edition, and “Who's Who in Science and Engineering”: 2011-2012, by in Marquis', USA, “2000 Outstanding Intellectuals of The 21st Century”, and “2000 Outstanding Scientists: 2010”, and nominated for “Top 100 Engineers: 2010”, by International Biographical Centre, Cambridge, England.. His areas of research include MATLAB/SIMULINK, Flexible AC Transmission Systems (FACTS), Modeling of Machines, Power Systems and FACTS, Controller Design, Power System Stability, Genetic Algorithm, Particle Swarm Optimization, Differential Evolution, Multi-objective Optimization, Economic Operation of Power System, Fuzzy Logic, Model Order Reduction, Distributed Generation and Wind Energy. Simanchala Padhy received B.E. degree in Applied Electronics & Instrumentation Engineering from R.E.C.(now NIT) Rourkela ,Orissa, India in the year 2001. He has also done the Post Diploma in Computer Application from U.C.P. Engineering School, Berhampur, Orissa, India in the year 1997 and Diploma in Electronics & Telcom. from U.C.P. Engineering School, Berhampur ,Orissa, India in the year 1995. Earlier, he has worked as a lecturer in V.I.T.A.M Engineering College, Parvathipuram, Andhrapradesh, India from 2002-2003. Presently he is working as an Engineering Assistant in Prasar Bharati, Broadcasting Corporation of India (Public Broadcaster of India) in HPT Berhampur, Orissa, India. C. Ardil is with National Academy of Aviation, AZ1045, Baku, Azerbaijan, Bina, 25th km, NAA 21