DSCformer: Non-Stationary Long-Time Prediction Methods Integrating Channel Spatial Attention

Time series fo recasting is crucial in many practical applications, and Transformer-based models have become cutting-edge tools for time series prediction. However, fo recasting tasks still face significant challenges when dealing with time series data characterized by noise and non-stationarity. This study proposes a new model based on the Autoformer architecture, called DSCformer, aiming to address long-term time series forecasting problems. Specifically, considering the simple linear mapping of the feed-forward module in the original model, we introduce a Channel-Space Attention Module (CSAM). This module allocates different attention to channels and spaces when extracting key information, enabling the model to suppress irrelevant noise information and accurately extract time series features. Additionally, in the self-attention mechanism, we introduce a De-stationary Attention module to extract non-stationary information. We applied this model to five real-world datasets, and experimental results indicate that compared to the original Autoformer, DSCformer reduces time series forecastin g errors by an average of 11.28%.