Study Of MG Time Series Forecast Based On Genetic Neural Network

The method of time series forecast is one of the important parts of research on science, economy, engineering, and so on. It is difficult for traditional method of time series forecast to predict the nonlinear system. The neural network has good nonlinear characteristic, and has offered a new approach for time series forecast. However, the neural network easily falls into local minimum, and weakly search the overall situation. The genetic algorithm (GA) has the ability of searching overall situation. The genetic neural network recombines the genetic algorithm's of seeking the superior overall situation and the neural network's nonlinear characteristic and rapid convergence. Mackey-Glass (MG) chaotic time series is one of benchmark problems in nonlinear time series forecast, which is representative. Therefore, this dissertation, from both theoretical and experimental perspectives, investigates several major problems on genetic neural network, and forecasts MG time series using genetic neural network. The main research achievements are as follows:A new mutation operator is constructed. Tournament selection is carried out with the produced three dimensional matrix which are represented the chromosome. Crossover is implemented using arithmetic crossover operator. Then unsymmetrical mutation is conducted using the new mutation operator which can expand the scope of chromosome gene value, at the same time, the chromosome with the highest fitness values are retained for each iteration. A lot of experiments are implemented to obtain the optimized initial weighs and bias.The back-propagation neural network is self-defined. The optimized weights and bias obtained by GA are used as initial weights and bias of the back-propagation (BP) neural network. Then, the neural network is trained by the variable study rate momentum of BP algorithm, the error goal is achieved by massive experiments.The numerical simulation to predict MG chaotic time series is conducted to validate the effectiveness of GABP (Genetic Algorithm Back-Propagation) neural network.