BSAformer: bidirectional sequence splitting aggregation attention mechanism for long term series forecasting

Abstract

Time series forecasting plays a crucial role across various sectors, including energy, transportation, meteorology, and epidemiology. However, existing models often struggle with capturing long-term dependencies and managing computational efficiency when handling complex and extensive time series data. To address these challenges, this paper introduces the BSAformer model, which leverages a unique combination of frequency-domain Sequence Progressive Split-Aggregation (SPSA) and Bidirectional Splitting-Agg Attention (BSAA) mechanisms. The SPSA module decomposes sequences into seasonal and trend components, enhancing the model’s ability to identify cyclical patterns, while the BSAA mechanism captures forward and backward dependencies, providing a comprehensive understanding of temporal dynamics. Extensive experiments conducted on seven benchmark datasets demonstrate the BSAformer model's superior performance, with notable improvements in accuracy and efficiency over state-of-the-art models. Specifically, the BSAformer achieves significant Mean Squared Error (MSE) reductions of 63.7% on the ECL dataset, 28.1% on the Traffic dataset, and 49.8% on the ILI dataset. These results validate the model’s robustness and its adaptability across diverse time series forecasting scenarios. The insights gained from this study contribute to the advancement of time series forecasting by providing a model that improves both accuracy and computational efficiency, especially in handling long-term dependencies and complex temporal patterns.