Time Series Prediction Based On Deep Belief Network

In recent years,with the rapid development of the Internet of Things system and big data technology,the collection of time series has become simple and fast,and the collected time series data has become huge in data volume,high in nonlinearity,and complex in data structure.Traditional time series analysis no longer satisfy the analysis and prediction of such complex data structures.In order to meet the prediction of highly nonlinear data structures,the research of new prediction models is particularly important.The emergence of deep learning brings hope to the prediction of highly nonlinear time series data.The Deep Belief Network(DBN)has powerful feature extraction ability and strong processing ability for nonlinear data structures.Compared with the traditional modeling method,the DBN can establish the model of the original dataa by using the "black box" through the extracted features.In recent years,the application of DBN in complex data structures such as image classification and speech recognition has been very mature,but it is still in the research and development stage in the application of time series prediction.In this paper,a new method is proposed for the modeling problem of time series prediction,and the time series prediction method is based on deep belief network.In this paper,the Echo State Network(ESN)is applied to the regression layer of the DBN to predict the time series.At the same time,DBN and ESN belong to the neural network,and the network parameters are more and have a greater impact on the network prediction results.Therefore,the Genetic Algorithm(GA)is used to optimize the network parameters.In the standard database(Mackey-Glass,MG)to test,compared with the traditional neural network,the prediction accuracy of the newly proposed GA-optimized DBN-ESN network is significantly improved,but in the case of excessive data,network training costs longer time.The prediction of short-term traffic flow is one of the applications of typical time series prediction.Time factor is a major factor affecting traffic flow.This test only considers the time factor of traffic flow,predicts short-term traffic flow,and applies the newly proposed network DBN-ESN to short-term traffic flow prediction.Comparison of predictions of other neural networks that exist.The research results show that the newly proposed network can highly fit the model of complex nonlinear data and has more accurate short-term prediction accuracy.When the amount of data is large,GA optimization network is time consuming.However,in the state where the traffic flow is close to saturation,people pay more attention to the accuracy of prediction,so the choice of the prediction model depends on the use environment.