Research On Multivariate Time Series Forecasting Based On Deep Learning

With the advent of the era of big data,it is easier for people to obtain data.Time series are also transformed from simple unitary time series to multivariate time series,and there are complex dynamic dependencies between time series.The classical statistical prediction model has been unable to grasp this relationship,so this thesis mainly uses deep learning model to capture the complex features between data.Multiple time series data often have dynamic periodic characteristics,trend and complex dynamic dependence.The traditional time series prediction model and deep neural network model can not accurately grasp these relationships.Therefore,the first model improved in this thesis is dynamic period neural networks(DPNet).In order to solve the problem that LSTNet can only remember a fixed period,DPNet introduces hidden attention mechanism(HAtt).HAtt module improves the ability of neural network to filter information.Its essence is to sum each memory information by weight,and learn the corresponding weight of each memory information through a large amount of data,the model can extract the key information that affects the trend of data,and then improve the prediction accuracy of the model.Through example analysis,the prediction effect of this model is better than VAR,LSTM,LSTNet and other models.The experimental results show that DPNet can effectively extract multiple features of periodic obvious data,so as to improve the accuracy of model prediction.When the volatility of multivariate time series data is severe,the traditional time series prediction model and deep learning network model can not accurately predict the volatile data.Therefore,the second improved model in this thesis is Dual-stage AttentionBased Recurrent Neural Network based on Wavelet Transform(WTDA-RNN).The model introduces wavelet decomposition module,DA-RNN module,LSTM module and wavelet reconstruction module.Wavelet decomposition is used to decompose the time series into multi-layer high frequency components and multi-layer low frequency components.Wavelet decomposition makes the model pay more attention to the trend characteristics and mutation characteristics of the data.LSTM module is mainly used to predict the lowest frequency coefficient data.DA-RNN module is mainly used to predict other high-frequency coefficient data,and wavelet reconstruction module reconstructs the predicted values of LSTM and DA-RNN to obtain the final predicted values.The results show that this model is better than VAR,LSTM,encoder decoder and DA-RNN.It shows that the improved WTDA-RNN model can effectively extract multiple features of volatile data,so as to improve the accuracy of model prediction.