Research On Time Series Forecasting Technology Based On Deep Neural Network

Time series prediction is one of the common problems in the field of data mining.The traditional time series prediction method establishes mathematical models to fit time series curves.However,the time series data generated in real life often has characteristics of chaos,irregularity and nonlinearity,so it is difficult to model accurately with simple mathematical models.Artificial neural network has the ability of self-learning and self-organization,and it can approximate the nonlinear model very well.Therefore,the application of artificial neural network in time series prediction problem can achieve good results.As a powerful machine learning tool developed in recent years,deep neural network has made great success in the fields of image,video and Natural Language Processing.However,in the field of time series prediction,the application of deep neural networks is still very few.In this paper,the deep neural networks are introduced to the problem of time series prediction,and the exploratory research is carried out.The main contents of this paper are as follows:(1)A prediction framework based on convolutional neural network is designed.The traditional time series prediction method is single and the prediction accuracy is not high.The time series prediction method based on machine learning can obtain high prediction accuracy,but it needs to extract the features manually,so the model is not common enough.In view of the above situation,a prediction framework based on convolutional neural network is proposed in this paper,which can automatically extract the characteristics of time series and get high prediction accuracy.The framework contains two volume layers.The first volume layer extracts the low order neighborhood characteristics of the time series,and the second convolution layers combine the low-level neighborhood features into high-order complex features.The max pooling layer is added after each convolution layer to reduce the model parameters and enhance the robustness of the model.The fully connected layer is added to the top of the framework to improve the expressive ability of the model.Experiments on the real dataset show that the proposed framework can predict the time series well.(2)An improved time series prediction framework based on recurrent neural network is proposed.The attention mechanism is introduced into the classic recurrent neural network framework,and the output vectors of each time step of the LSTM layer are weighted and summed to replace the output of the LSTM layer in the original framework.Compared with the original architecture,which only uses the output of the last time step,the improved framework can take the output of all time steps into consideration in a partial way and improve the prediction accuracy.The experimental results show that the improved framework can obtain higher prediction accuracy compared with the original framework.(3)A dual-channel time series prediction framework based on hybrid neural network is proposed.The dual-channel framework combines the convolutional neural network and the recurrent neural network.The convolutional neural network channel can extract deep features of time series by convolution and pooling operations,and the improved recurrent neural network channel can extract long sequence dependent features.Compared with the above two single channel frameworks,the prediction accuracy of the dual-channel framework on the whole test set is higher than that of the single channel framework.