The Identification, Modeling And Control Of Nonlinear System Based On SVM

The support vector machine regression(SVMR) based on Statistical Learning Theory(SLT) is a novel powerful tool of system identification and modeling. It is based on a complete theory and the idea of tradeoff between Emprical Risk Minimization (ERM) and Structural Risk Minimization (SRM).It provide high generalization ability and overcomes the overfitting problem experienced hy the other learning technique such as ANN. The training of SVM is done by quadratic programming possessing a global solution, which overcomes the problem of local minima suffered by classical neural network. An interesting and important property of SVM's solutions is that one obtains a sparse approximation, in the sense that its structure is naturally decided instead of decision based on experience of ANN. It could approach any nonlinear function. Initially, it was mainly used for pattern recognition, recent years, it was payed more attention to by control domain.This dissertation presents a SVM Regression method to identify and modeling nonlinear time-delay systems mainly about the system identification of system order and system delay. According to the sum of error and the number of support vector, evaluate if the system order and system delay is match. A simulation experiment is done to show the effect of this method.While the mechanism of this kind of system identification method is analysed based on the theory of SVM.Two application example is discussed to validate the ability of SVM for nonlinear identification. With he Superconducting Fault Current Limiter with Saturated Iron Core (SFCL) as the research object, we adopted SVM to identify the fault current to trigger the current limiter ,then the relationship between voltage and current of the varistor as the energy dissipation component of SFCL is identified by SVMR, that is used as a tool of simulation experiment for the energy dissipation of SFCL.This dissertation does some research about nonlinear internal model control (IMC) with SVMR. The SVMR is adopted to identify the internal model and the inverse model by the way of static offline. The influence to control precision of model error is analysed and then a linear compensation method is presented to decrease this kind of influence. By analyzing the relationship between regularization parameter C, insensitive parameterεand model error, control precision, it is pointed out that model error and model complexity are contradictory. Afterward an engineering design method of decreasing the model complexity , increasing the model error and the capability of model implementation, then linear compensation to the control error is presented.At the same time, a novel dynamic control method of adaptive internal model control to nonlinear time-variable system based on least square support vector machine (LS–SVM) is presented. LS-SVM is used to identify the internal model and inverse model. Because of the real-time performance of LS-SVM, according to the time window of the input and output datum of controlled object, system identification is running on line. So this kind of method could control nonlinear time-variable system. After that the relationship among the size of time window of input-output datum, model error and control precision is discussed, a determination method of the size of time window is given. A error guideline is used to decrease the number of adaptive identification, then to increase the real-time ability of the arithmetic.In this dissertation, we also use SVMR to study Model Predictive Control (MPC) to nonlinear system. We adopt SVMR to identify inverse model as an elicitation to find the dynamic optimal control law, that result in the decrease of search space and the increase of real-time ability of this kind of method. We call this kind of method as heuristic SVMR-MPC based on inverse model elicitation.Through simulation, the proposed methods were validated.