Research On Time Series Prediction Based On Deep Belief Network

A time series is a sequence ordered by a set of rules.Time series prediction is widely used in various fields,such as electric system,meteorological observation,financial analysis.And it plays a significant role in both theoretical researches and practical applications.With the developments of science and technology,the time series prediction methods are more diverse,from linear model to non-linear model,from single model to hybrid model,from traditional model to artificial intelligence model,the prediction accuracy is persistently improved,and the performance of the model is more perfect.However,due to the distinct characteristics of time series such as trend,seasonality,periodicity,the improvement of the accuracy of models has become a hot topic in time series prediction.Therefore,numbers of methods have been employed to enhance the performance of the model,such as the analysis of the dimensions of input variables,the adjustment of the structure and parameters of the model with the assistance of heuristic optimization algorithms,the noise reduction of the series with the data preprocessing technique.According to the the parameters and structure optimization and the data preprocessing of the model,two hybrid algorithms are proposed,which are the BP neural network model based on the restricted Boltzmann machine and the improved backtracking search evolutionary algorithm(RBM-BSASA-BP)as well as the deep belief network model based on the empirical wavelet transform(EWT-DBN).(1)The optimization of model structure and parameters.RBM-BSASA-BP first uses RBM to deal with data in a unsupervised rule,aiming at the abstraction of input relationship and the acquisition of the rough parameters with respect to the model.Moreover,the proposed method use BSASA algorithm to search global optimum parameters.Finally,the BP neural network is used as a predictor to make a prediction.The numerical experiments on data sets which include one artificial data set and two real data sets show that the model can effectively improve the prediction accuracy.(2)Data preprocessing.Data preprocessing technology is utilized to reduce or eliminate the influence of uncertain factors through series analyses before training.In order to extract the different patterns of time series,EWT-DBN uses the EWT algorithm to decompose the original sequence.Then,the similar components of the series are integrated for the purpose of capturing the data characteristics as well as shorten the training time.Finally,the DBN model is used to train and predict the integrated data.The experimental results of three real data sets involving traffic flow,load,and internet flow,show that the EWT-DBN model can obtain the more accurate forecasting result than other competitors.