Chaotic Time Series Prediction Based On Fuzzy Neural Network

This dissertation is on the basis of chaos dynamics and fuzzy system theory, the main research is focused on modeling and learning methods of fuzzy neural network that used in chaotic time series prediction, the concrete content and results are as follows:(1) A mixed BP learning method of fuzzy neural network is proposed. Using this method, the fuzzy rules can be drew and optimized from datum directly, the parameters of every fuzzy membership function and the weights of fuzzy rules can be adjusted with BP algorithm.(2) A new kind of encoding and decoding method of fuzzy rule is put forward while optimizing the fuzzy neural network with genetic algorithm, comparing with the binary encoding method and n-value encoding method, the length of the chromosomes can be respectively reduced by several hundred time and several time, and the time that the operations need can be shorten greatly.(3) The cause of lower precision of prediction is analyzed while employing improved genetic algorithm to optimize the fuzzy rules and parameters of fuzzy neural network, it is because that the more parameters need to be adjusting and the less generations is set. The calculating amount of one generation is so large that the whole necessary time is very long, so we have give a limited generations and the fuzzy network can't reach the optimal state.(4) The structure and parameters of the fuzzy neural network are adjusted by. Comparing with the above mix BP algorithm, the important information of fuzzy rules will not be lost, and the optimum fuzzy rules will be get unquestionably. The algorithm that combines BP algorithm with genetic algorithm can accelerate speed of convergence and prevent the defect of BP algorithm that easily to fall into local infinitesimal. Comparing with simple genetic algorithm, the necessary time of optimizing parameters is reduced greatly.(5) Using the fuzzy neural networks optimized with the above several algorithms to forecast the chaotic time series of Lorenz system, the results reveal that the fuzzy neural networks optimized with combining of GA and BP algorithm has the highest predicting precision, the networks optimized with mixed BP algorithm takes second place, the system that optimized with improved genetic algorithm gets the lowest predicting precision.