Research On Time Series Forecasting Method Based On Multi-attention Mechanism

The time series forecasting task studies the digital sequence of a phenomenon that changes with time,and finds its changing law.In actual life,there are many phenomena that contain time series,such as the trend of stocks,changes in the prices of goods in shopping malls,and changes in the flow of people in and out of train stations.Therefore,in-depth study of the time series issue is of great significance to the people of the national society.Among the solutions to the time series problem,there are mainly machine learning represented by the differential integrated moving average autoregressive model(ARIMA)and Back Propagation(BP)based on traditional statistics.Recurrent Neural Network(RNN)and variants commonly used in models and deep learning models.But they all have their own shortcomings.ARIMA has high requirements for the stability of time series data and poor prediction results for unstable time series data.The BP algorithm is easy to fall into the local minimum,and it is not good enough to choose the overall optimal scheme.Although today's deep learning model recurrent neural networks and their variants have been able to solve part of the gradient disappearance problem,these networks cannot remember long-term historical information,and the failure to join social emergencies will affect the final prediction accuracy.In order to solve the above problems,a time series prediction model N?ALSTM(News,Attention and Long Short Term Memory)based on multi-attention mechanism is proposed.Mainly from the two angles of news text and stock sequence data.The first is to address the problem of the lack of expression of social breaking news events in recurrent neural networks.This model uses Internet news with high real-time and high credibility as the basis of social events in time series prediction and analysis.A widely used word vector,selfattention mechanism,and Long Short Term Memory(LSTM)are used to construct the news public opinion analysis module in the model.In the time series prediction module,the encoder-decoder model and attention mechanism are used to output the structured data of the time series data and the news text features in the encoder for intermediate semantic output.The prediction results are more stable and accurate,avoiding time series Deceptive effects due to complex changes.Using N?ALSTM,time series forecasting methods that do not include news text analysis ALSTM,NS?TCL,TC?LSTM,HAN five algorithms were tested on five stock data sets,and found that N?ALSTM prediction accuracy is better than the other four algorithms.The prediction effect is slightly better.