Deep Learning with Dual Attention and RevINF for Time Series Prediction

Time series forecasting has long been a topic of widespread interest. With the advancement in computational power, the methods for time series prediction have gradually shifted from traditional machine learning approaches to deep learning methods, particularly with the widespread adoption of Long Short-Term Memory (LSTM) in time series problems. LSTM effectively capture dependencies between data points over time, yet they may not focus adequately on critical time points. To address this issue, we propose a dual attention mechanism that introduces attention mechanisms to better emphasize crucial nodes, thereby optimizing model performance. Furthermore, time series data are influenced by numerous factors, especially non-natural ones, which pose a significant challenge in time series forecasting. To tackle this problem, we introduce RevINF, a method that extracts non-stationary factors from time series data, predicts stationary data, and finally restores the non-stationarity of the data. This approach retains data information while enhancing the robustness of predictions, even in the presence of non-natural influences.