Decomposed Seasonal-Trend Network with Rotary Attention for Time Series Forecasting

Time series forecasting plays a crucial role in a wide range of realworld applications, such as energy management, weather forecast and financial analysis. However, entanglement of spatial and temporal patterns in time series limits forecasting accuracy of Transformer. To address the problem, decomposition can disentangle the trend and seasonal components of input series. We also notice that Rotary Position Embedding is effective for modeling seasonal signals, while linear architectures handle smooth trend signals well. Before modeling, down-sampling is used to capture multi-granularity information. Therefore, we propose Decomposed Seasonal-Trend Network with Rotary Attention (DSTN-RA) as a novel model for time series forecasting. It combines the decomposition with Rotary Attention and Channel Mixing modules to learn seasonal and trend signals separately. Experiments demonstrate that DSTN-RA outperforms leading baselines. Furthermore, ablation studies reveal differences in the information dependencies of seasonal and trend components, validating the decomposition method and confirming the specific contributions of each module. Codes will be available on GitHub.