Research And Application Of Time Series Analysis Based On ARIMA-LSTM Combining Model

Time series data often present high complexity,high correlation and long time dependence,and there are problems of insufficient extraction of series information,difficulty in capturing correlation and long time dependence in univariate time series forecasting and multivariate time series forecasting,and it is difficult for a single forecasting model to effectively extract deep features of the series.Therefore,it is of some practical significance and application value to study the time prediction forecasting method based on the combination model.In this paper,we analyze the current situation of time series forecasting research,combine the research results of deep learning,improve the problems in time series forecasting,and propose a time series forecasting method based on ARIMA-LSTM combined model.To address the problem of insufficient extraction of effective information of the series in univariate time series forecasting,the ARIMA model is introduced to forecast the long-term trend of univariate series,and the LSTM model is used to extract and fit the nonlinear information of the series to achieve sufficient extraction of effective information of the univariate series.To address the issue of insufficient extraction of effective information in single-variable time series prediction,the ARIMA model is introduced to predict the long-term trend of the single-variable sequence,and the LSTM model is utilized to extract and fit the nonlinear information of the sequence,achieving sufficient extraction of effective information in single-variable sequence.For the problem of difficulty in capturing correlation and long-term dependence in multi-variable time series prediction,this paper proposes the DA-LSTM model,which is based on the DA-RNN model and integrates whale optimization algorithm,attention mechanism,bidirectional LSTM model,and LSTM model.The input attention mechanism and bidirectional LSMT model are used in the encoder to improve the ability to capture sequence correlation,and the time attention mechanism and LSTM model are used in the decoder to avoid the problem of insufficient long-term dependence.The whale optimization algorithm is used to optimize the model parameters and reduce the impact of inappropriate model parameter settings on prediction accuracy.Then,this paper proposes a DA-LSTM model based on a two-stage attention mechanism and using a bidirectional LSTM model as an encoder and an LSTM model as a decoder,which enhances the ability to extract sequence information and avoids the problem of insufficient dependence for a long time.To reduce the impact of unreasonable parameter settings of the LSTM model on prediction accuracy and improve prediction accuracy,the whale optimization algorithm is used to optimize the number of hidden layers,learning rate,and iteration times of the bidirectional LSTM model and the LSTM model.The improved ARIMA and LSTM time series prediction methods are combined using parallel weighting and weighted least squares methods to construct the ARIMA-LSTM combination model.This combination model can capture both the long-term trend of the target sequence and the correlation between the target sequence and the influencing sequence,effectively avoiding potential problems in the time series prediction process and enhancing the ability to extract time series data features.Finally,the prediction performance and effectiveness of the ARIMA-LSTM combination model in different application scenarios are verified based on three datasets: SML2010,Appliances energy prediction,and Beijing Multi-Site Air-Quality.The experimental results show that the combination prediction model reduces the MAE and RMSE by2.22% or more compared to the other two common prediction models in different datasets,and by 4.01% or more compared to the two improved time series prediction methods in the paper.