Research On Self-Constructing Fuzzy Neural Network Control Algorithm Of AC Drive System

Due to its high level of simplicity, stability and reliability, conventional PID controlalgorithm is widely used in control objects with accurate mathematical models. Howeversatisfactory control performance can not be obtained if it is applied in control objectswithout accurate mathematical models or with nonlinear and uncertain characteristics.With the extension of problem domain and the development of simulation technology,applying intelligent control algorithm in control systems to deal with their nonlinearitiesand uncertainties has recently become a research focus in control field.In this thesis, a self-constructing fuzzy neural network (SCFNN) control algorithm isproposed based on synthesis of three core technologies in intelligent control, that is, fuzzycontrol, neural network and genetic algorithm. In this self-designing control algorithm,structure learning and parameter learning are performed online and simultaneously.Structure learning is based on partition of the input space. In order to implement structureself-regulating, it adjusts the number of fuzzy subsets, and accordingly adjusts the numberand distribution of membership functions, fuzzy logic rules, and connection relations andparameters of network through reinforcement learning. Parameter learning is based ongenetic algorithm, which possesses the ability of global search. It automatically adjustsnetwork parameters with absolute value of speed error as its fitness function andmaximization of reward reinforcement signal as its goal.In order to prove the effectiveness of the proposed SCFNN control algorithm andstudy on influence of network form on control performance, an AC asynchronous motor isintroduced as the control object in this study. Simulation model of the AC drive system isestablished based on MATLAB/Simulink platform and simulation research on SCFNNwith PD structure, forward and backward PI is performed separately. Plenty of simulationresults demonstrate the feasibility and effectiveness of the SCFNN control algorithm, andshow that SCFNN with PD structure has favorable dynamic characteristics but fails toeliminate the steady-state error, while SCFNN with forward or backward PI has lessfavorable dynamic characteristics but manages to eliminate the steady-state error. This thesis provide an effective method to deal with the adaptive control problems of the ACdrive system.