Research And Application Of Time Series Prediction Algorithm Based On Deep Learning

At present,data analysis research is still a hot spot in the industry.Relevant scholars carry out research from the perspectives of anomaly detection,intrusion detection and time series prediction in the key technologies of data analysis.However,existing studies have many limitations in time series forecasting,which do not take into account the computational consumption of complex and large scale data,ignore the time covariance drift of non-stationary data,and lack a real-time,accurate prediction model with strong generalization performance.In this thesis constructs a Fast?TS(Fast Time Series)forecasting model from the perspective of time series forecasting,using deep learning theory,to optimize the accuracy and computational effort of forecasting.The model is experimentally proven to be both accurate and computationally compact in the face of large and complex data.Based on this,the adaptive forecasting method A?Fast?TS(Adaptive Fast?TS)is proposed to improve the generalization of non-stationary time series forecasting.The main work of this thesis is as follows:(1)A sparse plus-attention mechanism method Fast?TS is proposed,which makes the time complexity of Transformer class time prediction model reduced to O(N),while improving information utilization by increasing the main vector its corresponding attention weights.The public datasets ETTh1,ETTh2,ETTm1 and the home-made dataset fighter demonstrate that the Fast?TS algorithm has a high level of prediction accuracy and low computational effort when facing long time series prediction problems with different prediction window lengths or different input window lengths.The RMSE of the Fast?TS algorithm decreases by 55.71% on average compared to the Reformer2020 algorithm for univariate prediction and by 79.89% on average compared to LSTNet for multivariate prediction;in the prediction experiments with different input window lengths,Fast?TS has the best results,with the difference between the prediction accuracy and Transformer being only 0.000,which is higher than that of LSTM* by about 17%.(2)The Fast?TS based adaptive prediction algorithm is proposed for the nonstationary time series Time Convariate Shift(TCS)problem.Finally,we compare Prophet,ARIMA,GRU,MMD-RNN,DANN-RNN,Light GBM,LSTNet,Transformer,STRIPE,ADARNN algorithms and Fast?TS algorithm on Air quality dataset by metrics RMSE and MAE to verify the A?Fast?TS algorithm's effectiveness.The experimental results show that both RMSE and MAE of A?Fast?TS are minimum,indicating the predictive superiority of the algorithm,and the model has excellent generalization with RMSE values reduced by 15.87% compared to Prophet on the Dongsi site and 13.14% compared to Prophet on the Tiantan site.(3)Design an intent prediction system for airborne target clusters to achieve future intent prediction based on the current perceived posture.The system includes data processing,parallel extrapolation,prediction,and visualization modules.The prediction module uses the previous Fast?TS algorithm to improve the problem of low accuracy and slow prediction of air target cluster intent prediction.