SPATH: A Transformer Spatiotemporal Prediction Framework with GAT-based Hierarchical Fusion Network

Effective temporal-spatial sequence prediction remains challenging due to the complex interactions between temporal patterns and spatial dependencies. Traditional approaches and recent Transformer-based models often struggle to capture these intricate relationships, either processing temporal and spatial features separately or using simple fusion mechanisms. To address these challenges, we propose a novel prediction framework based on GAT-enhanced Hierarchical Fusion Network (HFN). Our framework introduces three key innovations: (1) a hierarchical fusion architecture that progressively integrates temporal and spatial information through cross-attention and graph attention networks, (2) a novel aspect-based attention mechanism that dynamically weights feature dimensions based on their semantic importance, and (3) an optimized encoder-decoder structure that efficiently processes long sequences while maintaining prediction accuracy. Experiments on multiple benchmark datasets demonstrate that our framework consistently outperforms advanced Transformer variants including Reformer, Informer, and STTN, achieving significant performance improvements. The results validate the effectiveness of our hierarchical fusion approach in capturing complex temporal-spatial dependencies.