Identification and control of electrical circuits using neural networks

ABSTRACT Master Thesis IDENTIFICATION AND CONTROL OF ELECTRICAL CIRCUITS USING NEURAL NETWORKS Mehmet SAMAN Fırat University Institute of Scientific Department of Electrical-Electfonic 2003, Page:84 A classical and modem control theory are based upon linear models. However, the feet that most practical systems are nonlinear constitutes an enormous difficulty for doing successfully many applications. Using artificial neural networks (ANNs) that have naturally nonlinear mapping capability is an alternative and powerful solution to nonlinear control and identification problems. Different architectures of ANNs have been used in the literature. In this thesis, three different ANNs architectures are covered: feedforward neural networks (FNNs), radial basis function neural network (RBFNN), and modular neural networks (MNNs). Although FNN gives good performances in system identification, long training time for its learning İs required. The learning algorithms are used to improve their identification performances. Levenberg Marquard, Gauss-Newton, and Conjugate Gradient here are taking into consideration and discussed by the simulation results. When MNNs are compared with FNNs, they provides more performances than FNNs for only piece-wise linear system in term of less training time. Because MNNs combines different the learning algorithms and ANNs architectures. On the other hand, RBFNN trains in very small time and their performances changes according to their linear weights and radial basis functions parameters.Based on these learning methods and identification capability nonlinear system of ANNs, inverse control, internal model based control, optimal control and predictive control methodologies are applied on the example electrical circuits and excessive simulation results are discussed in this thesis. Keywords: ANN, FNN, MNN, RBFNN, Neuro-Control, Guasi-Nevton, Backpropagation, Levenberg-Marquardt, Conjugate-Gradient.