YSA'lı rotor akışı gözlemcilik vektör denetimi

Abstract

Bu tezde sincap kafesli asenkron makinanın gerilim aradevreli rotor akısı yönlendirilmiş doğrudan vektör denetiminin simülasyonu gerçekleştirilmiştir. Akı modelinin bir bölümünde ileri yönde çok katmanlı algılayıcılı bir YSA kullanılmıştır. Makina modelinde ise, isO, isD, imrD, imrQ durum değişkenleri olarak seçilmiş ve matematiksel model 4. mertebeden Runge-Kutta sayısal integrasyon yöntemiyle çözülmüştür. Model oluşumu sırasında uzay fazörlerinden yararlanılmıştır. Dik eksenli asenkron makinanın simülasyonu MATLAB programlama dilinde yazılmıştır. Tezde, asenkron makinanın DA makinasına vektör denetimi yardımıyla benzetilmesi anlatılırken asenkron makinanın rotor akısı yönlendirilmiş vektör denetiminin simülasyonu da yapılmıştır. Gerilim aradevreli tipinde olduğu için kullanılan kontrolör sayısı da fazla olmuştur. 5 adet kontrolörün 10 adet katsayısının deneme yanılma yöntemiyle seçilmesinden ötürü sistemin optimum çalıştığını iddia etmek doğru olmayacaktır. Uygun parametre seçimi kararlık, kararlı hal hatalarının sıfıra gitmesi, ve cevaptaki aşımların az olması gibi olumlu sonuçlar doğurur. Yapay sinir ağlan hakkında genel bir bilgi verildikten sonra çok katmanlı ağlar incelenmiştir. Geriye yayılım algoritmasının çıkarımı, hem çıkış katmanındaki hem de saklı katmandaki bir nöron için yapılmıştır. Hatanın geriye doğru yayılması gösterilmiştir. Ardından yapay sinir ağları ile dinamik doğrusal olmayan sistem tanıma hakkında bilgi verilirken bir adet örnek simüle edilmiştir. Gerilim aradevreli vektör denetiminin simülasyonunda yapay sinir ağlan akı modelinde kullanılmıştır. Akı modelini oluşturan iki denklemden sadece rotor mıknatıslama akımının genliğini veren denklem normalize edilerek ağa öğretilmiştir. Eğitimde momentum sabiti ve öğrenme katsayısı uygun seçilmiştir. Eğitilen ağın da katılmasıyla oluşturulan vektör denetimi istenilen sonucu verebilmektedir. Çizdirilen şekillerle yapay sinir ağlarının vektör denetiminde kullanılabileceği kanıtlanmıştır.

In this thesis various simulations are given for obtaining direct vector control of a cage induction machine with a voltage-fed inverter under rotor-flux-orientation where a part of the flux model is realised with a feedforward artificial neural network. The whole induction machine model is discretized with fourth order Runge-Kutta numerical integration method. However Euler method is also used for mechanical part of the system where friction constant is omitted. The space phasor theory has been used to derive the integral part of the electromechanical side of the induction machine where stator windings are uniformly distributed in order to obtain a sinusoidal MMF around the rotor periphery. Therefore machine can be classified as a symmetrical machine. Under these assumptions, the induction machine models are given The last one which is deduced by the help of space phasor theory is expressed as, Rs+pLs "COiLs PLm -(öjLn CÖlLs Rs + PLs COxL,» PLm PLn -SG)lLn Rr + PL -SÛ)lL s^Ln pL» scOxL Rx+pL. 1«. L (2.44) where stator and rotor quantities are given in quadrature axes. Using the isQ, iso, Wd, imrQ and Or as state variables induction machine model is simulated using MATLAB programming language. Electromagnetic torque versus angular speed diagram is depicted as below. In Chapter 3 vector control of induction machine is summarized. The comparison between DC machines and induction machines - in electromagnetic torque production - is given. The necessary steps, which should be taken into account to achieve decoupling of the stator current to flux-producing and torque-producing components, are explained in detail. The basis of vector control of AC machines is explained by the analysis of magnetizing-flux-oriented vector control of cage (short- circuit) induction machine. The difference between the direct and indirect vector control of AC machines is emphasized and additionally direct method using a flux model is thoroughly studied. Because of the selection made for the inverter type, the stator voltage equations are not dropped from the vector control part of the drive system. 50 100 150 200 Açısal hız [rad/s] Figure 2.9 Electromagnetic torque vs. angular speed Therefore a decoupling circuit which decouples (ending the interaction between two axes) the is* and the i^, from each other. If the converter is on the rotor side, the monitored rotor currents will ease the implementation. On the contrary, if the machine is supplied by a current-controlled PWM with fast current control loops, the stator can accurately follow the reference currents while the stator equations are not included in the drive's model. In the voltage-fed rotor-flux-oriented direct vector control of induction machine, the independent control of the direct and quadrature axis of the stator currents is required. This can only be achieved if the stator voltage equations are decoupled and two stator components are indirectly controlled by controlling the terminal voltages of the machine. In the model used, a FG (Function Generator ) is employed in order to cope with the inconvenient situation above base speed. By the help of the FG, field weakening is achieved. In the FG, below base speed a constant maximal value is obtained and above base speed this value is reduced in inverse proportion to the rotor speed. FG is implemented in MATLAB code as below (for imrnonrinai=6. 1 A), FUNCTION GENERATOR if abs(wr)>184 & abs(wr)<265 imref=6. l-abs(wr)/30. 16+6. 1; elseif abs(wr)<=l 84 imref=6.1; else imref=0; end XIV Another important point in the simulation is the PI controllers. 5 PI controllers are employed where one is optional. The appropriate tuning of the controller parameters makes the system, stable and steady state error free. The flux model, which will be partially replaced by the artificial neural networks, is defined from the rotor equations. T,%U|rjN. (318) dt %^^+^î (3.19) ^ Tr|i"| The whole equations governing the vector control scheme (including PI controllers, Function generator for field-weakening region, decoupling circuit, flux model, torque and reference voltage limiters etc.), vector transformers and the machine are analyzed through a simulation program written in MATLAB language. As can be seen from the results, the system works even under sudden load changes within a wide reference speed range. In Chapter 4, artificial neural networks are explained. The types of neurons, the basic computational element in the networks, and the network architectures are given. Different types of training methods are mentioned where the emphasis is put on the error-back propagation algorithm. Also the perceptrons,the simplest form of a neural network used for the classification of a special type of patterns said to be linearly separable, are analyzed. Multilayer perceptrons with their forward and backward phases are explained. 6j(n) = cp'k(vk(n)) £8k(n) wkj(n) (4.27) k Awji(n) = Ti8j(n)yi(n) (4.21) Neuron J' s, which is in a hidden layer, weight change for each step is given above. Other types of neural networks are also mentioned. The derivation of BP (Back Propagation) algorithm is made for neurons both in the output layer and in the hidden layers. Error Back Propagation algorithm is shown in diagrammatic form. In this chapter system identification via neural networks are also discussed. Two types of identification methods, with their advantages and disadvantages, are explained. Series-parallel identification method is preferred for its stability. A simulation which identifies a simple system is made. yp(k+l) = 0.3yp(k)+0.6yp(k-l)+f[u(k)] (4.3 1) fiu) = 0.6sin(7cu) + 0.3sin(37tu) + 0. lsin(57cu) (4.32) XV The Equation (4.32) is learned to the network and the model is compared with the system's response. Figure 4.9 Diagrammatic scheme of Back Propagation xvi In Chapter 5, a program which trains the network to learn the magnitude of the rotor magnetizing current is written. The equation derived for training is given below: Lr (* + !)",*t (*) = *t W, '« (*) (6l) h T T K ' } This equation is the discretized form of the Equation (3.18). However the output of the neural network is a squashed function. Therefore the outputs - including the inputs - must be kept in the range of [-1,+1] by a normalization process. The necessary normalization constants, K*, must be greater than the maximum value that can be found in the associated variable. After the normalization process the equation of the system which is to be identified is trained. The stated equation is as follows, Wn(k+1)= fl-Aİi^Ck^ ^L/.(t) (6.2) After training is accomplished, the artificial neural network is embedded into the previously written vector control simulation -with the computed (constant) weights. The vector control of induction machine responded in a similar manner when the flux model is used. Finally, the advantages and the shortcomings of artificial neural networks in system identification while putting emphasis on field-oriented control are thoroughly discussed.