Bi-TimeSSM: A Bidirectional Time-aware State-Space Model for Long-Term Time-Series Forecasting

Long-term time-series forecasting (LTSF) remains a challenging task due to the need to jointly capture complex temporal dependencies and intricate inter-variable interactions. Structured State Space Models (SSMs), such as Mamba and S4, have shown great promise in modeling long-term dependencies with high efficiency. However, existing SSM-based approaches often fall short in two aspects: limited bidirectional modeling capability and static channel interaction mechanisms. Some methods employ parallel Mamba models to process original and transposed sequences separately, but their predefined fusion strategies lack adaptability to diverse temporal patterns. Others enable bidirectional processing yet struggle with effectively capturing long-term dependencies in high-dimensional multivariate settings. To address these limitations, we propose Bi-TimeSSM, a novel bidirectional SSM-based architecture designed to enhance both temporal and inter-channel representation learning. Bi-TimeSSM incorporates advanced state-space modules for scalable multiscale modeling and introduces a dynamic channel interaction mechanism that adaptively determines whether and how to fuse information across variable channels. Extensive experiments on multiple benchmark datasets demonstrate that Bi-TimeSSM consistently outperforms existing models, offering a robust and flexible solution for multivariate LTSF tasks.