TSM-ATTN: A Hybrid Time Series Forecasting Model Combining Attention Mechanism with Mamba

Time series forecasting requires addressing both short-term dynamics and long-term dependencies. Existing methods often prioritize long-term dependencies, overlooking short-term variations. While self-attention captures long-term dependencies, it struggles with short-term dynamics and suffers from quadratic complexity. Mamba efficiently handles short-term dynamics and local features but struggles with long-term dependencies and global interactions. This paper proposes a Hybrid Gated Attention module, which utilizes Multi-Head Self-Attention for feature extraction, and employs Mamba to select key dependencies for computational efficiency. A Multi-Layer Perceptron processes global and local features to enhance performance. Based on this module, the TSM-Attn model with a dual-branch structure improves accuracy and efficiency by focusing on global trends and local dependencies. Experiments on benchmark datasets show significant improvement.