Research On Analysis And Prediction Method Of Time-Series Data Based On Deep Learning

In today’s society,as data becomes complex and huge,there are more and more hidden information,values,and laws in it.As a typical type of big data,time-series data often has the characteristics of huge data volume,close time points,and many related variables.Now time series data often have multiple time-related variables,and the change of each variable depends not only on its historical data value but also on the influence of other related variables on it.The complexity of time-series data makes the information contained in it more valuable,but it also brings great challenges to the analysis and prediction of time series data.How to analyze and predict more accurate and valuable information from complex and huge time series data has become the focus of current data analysis researchers.Currently,time-series data analysis and prediction methods can be classified into three categories:basic periodic laws,machine learning,and deep learning.Based on the existing problems in time series data analysis and forecasting methods,this thesis further studies the analysis and forecast of hidden information in time-series data,and the main research work is as follows:（1）Aiming at the multivariate time series data sets collected by enterprise sensors,an AR-LSTM analysis and prediction model based on ARIMA and LSTM is proposed（Prediction Of SO₂Concentration Based On AR-LSTM Neural Network）.We first processed the outliers and missing values in the data sets;secondly,we performed the data set labeling and feature extraction;then we analyzed and predicted using our proposed analytical prediction model;then,we used R²,RMSE,and MAE as the evaluation indexes of the model prediction performance,and we compared our proposed analytical prediction model with other analytical prediction models on the two datasets respectively and the results demonstrate the effectiveness of the analytical prediction model.Finally,the AR-LSTM analysis prediction model is used to predict the data in the next few weeks.（2）Aiming at the complex interdependence between variables in multivariate time series data,an ATT-LSTM analysis and prediction model based on attention mechanism and LSTM is proposed.（Multivariate Time Series Data Prediction Based on ATT-LSTM Network）.This analytical prediction model applies the attention mechanism to the LSTM,which enables the analytical prediction model to screen the mutual influence information in multivariate data when analyzing and predicting multivariate time-series data,making up for the weakness of the LSTM in handling multivariate data,and greatly improving the accuracy of the network in predicting multivariate time-series data.The analytical prediction model proposed in this work was tested against other models on two datasets,NASDAQ-100 and Beijing PM2.5 concentration,using MAE and RMSE as the evaluation index of the model.The results demonstrate the effectiveness of the ATT-LSTM analytical prediction model.In summary,this thesis proposes two deep learning-based analytical prediction models based on the time-series data set collected by an enterprise sensor and the network public time-series data set,the time series data analysis and prediction work is carried out.The experimental results of both models demonstrate the effectiveness of the proposed analytical forecasting model in analyzing and forecasting time series data.