Application Of Prediction Of Time Series With Self-organizing Neural Networks And Their Hybrid Models

This paper introduces the basic concept and main features of time series, and analyzes time series through artificial neural networks and their hybrid models. Firstly, by using historical data of time series to train the neural networks we can describe the rules and the development trend of time series. And then we may forecast the target values of new data; and finally we will compare these values to the real target values of the data.In this paper, the neural network models selected are BP neural network which can approximate any continuous function with arbitrary precision, and self-organizing neural networks which have a very good performance in data classification. Then the self-organization combining with difference models are constructed based on the self-organizing neural networks.And the weights and the classification number of the models are improved.By comparison the author find out a better neural network structure. Combining the self organizing neural networks and BP neural networks, a better model is selected through comparing the prediction errors of different classification and number of neurons in hidden layer. The self-organizing neural network and regression model are combined, and improved with initial neural network weights and learning rate. Through the comparison of different classification number a model is found to have little error. The BP neural network is selected as a comparison model to evaluate the effect of the new models. Time series model tested in this paper are the Mackey-Glass time series model and stock index time series model. All simulation, prediction and comparison are carried out in the matlab tool.In the tests on prediction of Mackey-Glass time series,the smallest errors of BP neural network is0.0010, while self-organizing models combining with difference models0.3691, self-organizing combining self-regression neural network0.0008, and SOMBP neural networks0.0081. In the forecast of stock index time series, the errors of the above models are respectively0.0174,0.0081,0.0135and0.0381.The comparation of the results shows that in the Mackey-Glass time series, the error of is0.0002little than that of BP models,and in stock index time series,the errors of self-organizing models combining difference model is0.0093little than that of BP models.