Interior Permanent Magnet Synchronous Motor (IPMSM) Adaptive Genetic Parameter Estimation

Transcription

1 Interior Permanent Magnet Synchronous Motor (IPMSM) Aaptive Genetic Parameter Estimation Java Rezaie, Mehi Gholami, Reza Firouzi, Tohi Alizaeh, Karim Salashoor Abstract - Interior permanent magnet synchronous motors (IPMSMs) are receiving increase attention for high performance rive applications because of their high power ensity, high efficiency an flux weakening capability However, their high efficiency characteristic is influence by applie control strategies Thus much effort has been irecte towars the efficiency optimization of the IPMSM by minimizing motor copper loss This loss is influence by machine parameters Thus online estimation of these parameters is essential On the other han Genetic Algorithm (GA) is a tool for optimization an can be use for solving some problems that can be formulate in those forms that this algorithm can hanle it In this paper we formulate nonlinear state euation of this motor in such form that we can use GA for estimating the unknown parameters Simulation results show that the estimate parameters converge to correct values after several iterations Inex Terms Genetic algorithm; parameter estimation; nonlinear system; IPMSM; copper loss; efficiency I INTRODUCTION If all moel parameters of a linear/nonlinear system are known we can use KF/EKF for state estimation In some nonlinear cases EKF iverges In orer to solving this problem we must use the other tools for estimation One of these tools is Genetic Algorithm (GA) that uses the principles of evolution, natural selection, an genetics to offer a metho for parallel search of complex spaces This paper escribes a GA that can perform on-line aaptive parameter estimation for nonlinear systems (IPMSM in this paper) Manuscript receive July 22, 2007 This work was supporte in part by the International Association of Engineers (IAENG), engineering letters an CSA (Cambrige Scientific Abstracts) J Rezaie, Control an Intelligent Processing Center of Excellence, University of Tehran, Iran ( jrezaie@eceutacir) M Gholami, Automation an Instrumentation Department, Petroleum University of Technology, Iran (phone: ; metigholami@ gmailcom) N Firouzi, Automation an Instrumentation Department, Petroleum University of Technology, Iran ( nafirouzi@yahoocom) T Alizaeh, Automation an Instrumentation Department, Petroleum Uuniversity of Technology, Iran ( yahoocom) K Salashoor, Automation an Instrumentation Department, Petroleum University of Ttechnology, Iran ( yahoocom) It is a stochastic process which attempts to fin an optimal solution for a problem by using techniues that are base on Menel's genetic inheritance concept an Darwin's theory of evolution an survival of the fittest This paper introuces a moifie approach to aaptive parameter an state estimation where a genetic algorithm is use to estimate the plant parameters an hence to obtain an estimate of the state Parameter an state estimation problem formulate in such a form which we can use GA for solving the problem Finally simulate the problem an results will show The moel that we use in this paper is the state space moel of Interior Permanent Magnet Synchronous motor (IPMSM) which is nonlinear The remainer of this paper is organize as follows A base 10 genetic algorithm is presente in Section 2 IPMSM moel is iscusse in Section 3 In Section 4 formulate nonlinear parameter an state estimation problem formulate in such a form which we can use GA for solving the problem Finally simulate the problem an results will show II A BASE 10 GA In orer for the GA to fin the optimal solution to a particular problem, the parameters that comprise a solution must be encoe into a form upon which the GA can operate To borrow a term from biology an genetics, any set of parameters which may be a solution to the given problem is calle a chromosome, an the iniviual parameters in that possible solution are calle traits Since the GA will most likely be implemente on a igital computer, each trait must be encoe with a finite number of igits (calle genes) The more genes in a given trait (or in a chromosome), the longer the GA will take for encoing an ecoing purposes an in other operations, so a reasonable length shoul be chosen The entire set of chromosomes (that is, the entire set of caniate solutions to the given problem) upon which the GA will operate is calle a population [1] To evolve the best solution caniate (or chromosome), the GA employs the genetic operators of selection, crossover, an mutation for manipulating the chromosomes in a population A

2 brief escription of these operators follows The GA uses these operators to combine the chromosomes of the population in ifferent arrangements, seeking a chromosome that maximizes some user-efine objective function (calle the fitness function) This combination of the chromosomes results in a new population (that is, the next generation) The GA operates repetitively, with the iea that, on average, the members of the population efining the current generation shoul be as goo (or better) at maximizing the fitness function than those of the previous generation The most fit member of the current generation (that is, the member with the highest fitness function result, or "fitness value'') at the time the GA terminates is often taken to be the solution of the GA optimization problem [1] The first genetic operator use by the GA for creating a new generation is selection To create two new chromosomes (or chilren) two chromosomes must be selecte from the current generation as parents As is seen in nature, those members of the population most likely to have a chance at reproucing are the members that are the most fit The techniue use in [2] for selection uses a "roulette wheel'' approach [1] Consier a roulette wheel that is partitione accoring to the fitness of each chromosome The fitter chromosomes occupy a greater portion of the wheel an are more likely to be selecte for reprouction In [3], a selection metho is chosen so that a given segment of the population corresponing to the most fit members (that is, the D most fit members) are automatically selecte for reproucing Therefore, the least fit members have no chance of contributing any genetic material to the next generation Of the D most fit members of the current population, parents are ranomly chosen, with eual probability The latter metho of selection is use in this stuy [1] Once two parents have been selecte, the crossover operation is use Crossover mimics natural genetics (that is, "inheritance'') in that it causes the exchange of genetic material between the parent chromosomes, resulting in two new chromosomes Given the two parent chromosomes, crossover occurs with a user efine probability p c Accoring to [1], if crossover occurs, a ranomly chosen "cross site'' is etermine All genes from the cross site to the en of the chromosome are switche between the parent chromosomes, an the chilren are create Another approach to crossover, one that is use in [3] an in this stuy, is that crossover occurs exactly once (that is, p c = 1) for every trait, with the cross site within that trait chosen ranomly That is, all genes between the cross site an the en of the trait are exchange between the parent chromosomes Crossover helps to seek for other solutions near to solutions that appear to be goo [1] After the chilren have been create, each chil is subjecte to the mutation operator Mutation occurs on a gene-by-gene basis, each gene mutating with probability pm If mutation oes occur, the gene that is to mutate will be replace by a ranomly chosen allele (in this case, a ranomly chosen value between 0 an 9) The mutation operator helps the GA avoi a local solution to the optimization problem If all of the members of a population shoul happen to converge to some local optimum, the mutation operator allows the possibility that a chromosome coul be pulle away from that local optimum, improving the chances of fining the global optimum However, since a high mutation rate results in a ranom walk through the GA search space, pm shoul be chosen to be somewhat small [4] We have foun, however, that in some instances in real-time systems, we nee a slightly higher mutation rate This is the case since the fitness function epens on the ynamically changing state of a system, so the locality of an optimum is time-epenent an we must ensure that the GA is reaily capable of exploring new opportunities for maximizing the time-varying fitness functions [1] If a chromosome is generate by crossover an mutation, it is possible that one or more of its traits will lie outsie the allowable range(s) If this occurs, each trait that is out of range shoul be replace with a ranomly selecte trait that oes fall within the allowable range [1] In aition to selection, crossover, an mutation, a fourth operator can be use by the GA This operator, known as elitism, causes the single most fit chromosome of a population to survive, unisturbe, in the next generation The motivation behin elitism is that after some sufficiently small amount of time, a caniate solution may be foun to be close to the optimal solution To allow manipulation of this caniate solution woul risk unsatisfactory performance by the GA Therefore, with elitism, the fitness of a population (seen as the fitness of the best member of the population) shoul be a non ecreasing function from one generation to the next If elitism is selecte, the most fit member of the current generation is automatically chosen to be a member of the next generation The remaining members are generate by selection, crossover, an mutation [5] Notice, also, that this allows us to raise the mutation probability since we know that we have a goo solution available In [3] as well as in this stuy, elitism can involve more than just one member That is, a certain number r (possibly more than one) of the most fit members will survive in the next generation without manipulation by crossover or mutation If the most fit member woul point to a local optimum in the GA search space, but a slightly less fit member points to the global optimum, they might both survive in the next generation with this new form of elitism [1] To initialize the GA, a chromosome length must be chosen, along with the length an position of each trait on the chromosome The allowable range for each trait must also be specifie The population size (enote N) must be specifie, along with the metho of generating the first population The iniviual members may be ranomly generate, or they may be initialize to some set of ``best guesses'' In this stuy, a ranomly generate initial population is always use In aition, pc an pm must be specifie After this initialization, the GA can operate freely to solve its optimization problem [1]

3 III IPMSM MODEL Uner assumption of linearity of the magnetic circuit a set of wiely use moel for IPMSM base on a synchronous - reference frame, incluing copper is presente in figure 1 an figure 2 [6,7]: The moel euations are given as follows: i L s 1 = R i + ω i e + v t L L L (1) i Rs L 1 1 = i ωei λmωe + v t L L L L (2) 3P Te = λmi + ( L L) ii 2 (3) ωm J = Te TL Bωm t (4) ωe = Pωm (5) T e Pc = R s i P λm + ( L L) i (6) Where: i i - an -axis components of armature current, v v - an -axis components of terminal voltage, ωe Electrical angular spee, Rs Armature wining resistance, λ Magnetic flux linkage m λ λ -axis an -axis components of armature selfinuctance, TL Electro magnetic torue, P Number of pole pairs, Pc Copper loss All states of this machine are measurable an can be use for parameter estimation IV PROBLEM FORMULATION The above euation can be written in the following form: ( X ( t)) = F( X ( t), u( t), t), t Y () t = G( X (), t u(),) t t Discrete version of this euation with sampling time Ts is: X ( k + 1) = H( X ( k), u( k), k), Y ( k) = Z ( X ( k), u( k), k) In IPM synchronous machine all state ( I, I an ω m) are accessible an we can write the estimator moel in the form of system moel but with unknown parameters The cost function that must be minimize in the GA is the sum of suare estimate values of state subtracte from real values of state If the parameter chose such that this cost function minimize then the estimate value of system parameters is goo other than we must run GA such that this cost function minimize or near minimization region From the above we efine the following cost function: (7) (8) 1 ˆ 2 J ( Rs, L, L, λm) = {( I ( k) I ( k)) + 2 ( I ( k) Iˆ ( k)) ( ( k) ( k)) } ωm ˆ ωm t k = t n (9) where: Figure 1: -axis euivalent circuit Iˆ, ˆ I an ˆm ω are estimate values of I, I an ω m, k is current time an n is length of ata use in cost function Figure 2: -axis euivalent circuit V MODIFIED GA FOR ADAPTIVE PARAMETER ESTIMATION First we convert iscrete, continues system euations then using above mentione cost function in GA In this example, GA has (i)population size is 50, (ii)generations per each iteration is 10, (iii)mutation function is uniform with rate eual

4 to 02, (iv)crossover function is intermeiate with rate eual to 05, (v)selection function is roulette while Because sometimes the algorithm converge to some values (local optimum) that is not correct we moifie this algorithm with checke this that if the current time is grater than some value(such as 30) an cost value ( J () ) is grater than some value (such as 05,this means that the converge parameters is not eual or near eual of the real values) an current values of estimate parameter subtracte from past estimate parameters is smaller than some value (such as 001, this means the parameters was converge to some value) then we increase the mutation rate to some value grater than current value to exit this point else algorithm runs with past mutation rate From the estimation theory we know that if the input signal was not Persistent Exciting (PE) of sufficient orer the estimation was not converge to true values so we chose input signals ( V an V ) such that this conition satisfy Figure 4: -axis actual an estimate current VI SIMULATION RESULTS The results show that this algorithm estimates the parameters of nonlinear moel with appropriate precision an time Simulation results show that if the range of parameters where use in GA is small then convergence is faster an estimation is more accurate than using wier range for parameters VII CONCLUSION In this paper first we present the state euation an moel of IPMSM We see that this euation is nonlinear in nature Figure 5: actual an estimate spee Figure 6: actual an estimate R s Figure 3: -axis actual an estimate current

5 Figure 7: actual an estimate L iteration GA minimize this cost function an estimate parameters at each iteration After this step we introuce a moification in online parameter estimation using GA in orer to increase precision of the estimation an avoiance wrong convergence Finally we simulate that algorithm using MATLAB For a typical IPMSM those parameters of this motor will present at follow Simulation results shows that by implementing this algorithm we can estimate parameters of this motor aaptively an states converge to true values after several iterations This estimator is an alternative for parameter (or state) estimation for linear an nonlinear systems similar to KF an EKF In some nonlinear case that EKF iverge we can use this algorithm for parameter estimation APPENDIX Machine Specifications Rate spee, rpm 1800 Rate torue, Nm 396 P, No of pole pairs 2 Rs, Ω 193 L, L, mh 4244,7957 λ m,wb 0314 J, rotor inertia constant, 0003 Nm/ra/sec Figure 8: actual an estimate Figure 9: actual an estimate L λ m REFERENCES [1] J R Gremling an K M Passino, Genetic aaptive state estimation, Engineering Applications of Artificial Intelligence, 13, , 2000 [2] D E Golberg, Genetic Algorithms in Search, Optimization, an Machine Learning, Aison-Wesley, 1989 [3] L Yao an W A Sethares, Nonlinear parameter estimation via the genetic algorithm, IEEE Transactions on Signal Processing, 42, , 1994 [4] S H Thilagar an R G Srihara, Parameter estimation of three-wining transformers using genetic algorithm, Engineering Applications of Artificial Intelligence, 15, , 2002 [5] C S Ong, J J Huang an G H Tzeng, Moel ientification of ARIMA family using genetic algorithms, Applie Mathematics an Computation, 164, , 2005 [6] S Morimoto, Y Takea an T Hirasa, Loss minimization control of permanent magnet synchronous motors, IEEE Trans In Electron, 41, ,1994 [7] S Vaez an V I John, Minimum Loss Operation of PM Motor Drives, In Proc Of Canaian Conf on Elec & computer Eng, Montreal, pp , 1995 Then escribe a base 10 GA an formulate the parameter estimation in such away that can be solve with GA Because all state of this machine is measurable we use this measurement an estimate values of this states then ifference of real values from estimate values an use this value in cost function we want to minimize, then at each