A study on the application of ARIMA and MLA-LSTM hybrid model in time series forecasting

In this paper, we explore the application of a hybrid model combining ARIMA and LSTM with multiple latent attention (MLA) in time series forecasting.ARIMA model is good at dealing with linear relationships, transforming non-stationary series into stationary ones and modeling them through techniques such as differencing, auto regression, and moving averages, but is limited in dealing with non-linear relationships.LSTM, as a deep learning method, effectively captures long-term dependence and nonlinear relationships in time series, but it is not as efficient as ARIMA in dealing with linear relationships. Therefore, combining the two to construct a hybrid model, first extracting the linear patterns and calculating the residuals with ARIMA, then inputting the residuals into LSTM to capture the nonlinear patterns, and finally adding the two parts of the prediction results to obtain the final prediction value, which can make full use of their respective advantages and improve the prediction accuracy. And so it was concluded, Experiments show that the hybrid model is better than a single model in terms of prediction accuracy and robustness, and exhibits a broad application prospect.