HAM-MSTF: A hybrid attention-based model for multi-task spatio-temporal forecasting.

Abstract

Spatiotemporal forecasting plays a crucial role for optimal operations and decision making in various fields, including industries, transportation, healthcare and energy. However, spatiotemporal models are typically designed for a single task, failing to capture the inherent complexity and interrelationships among multiple tasks in real-world systems. To address this issue, we propose a hybrid attention-based model for multi-task spatiotemporal forecasting (HAM-MSTF) with the help of dual attention mechanisms. Unlike conventional single-task spatiotemporal forecasting, spatiotemporal multi-task learning jointly predicts multiple target variables over the same spatial system and time horizon through shared spatiotemporal representations and task-specific adaptation. Firstly, it extracts spatial features into a high-dimensional latent space to enrich spatial information by using the graph attention mechanism based on the multi-gate mixture-of-experts framework. Then the proposed method integrates both global and local temporal dynamics, to make full use of fast feature weights to enhance temporal attention and achieve independent spatiotemporal feature fusion for each task. This architecture not only effectively captures spatial and temporal patterns but also accounts for inter-task differences, thereby improving prediction performance across diverse tasks. Finally, a series of evaluations are conducted on one synthetic dataset (concentrated suspension two-phase flow) and two widely-used benchmark datasets (PEMS04, PEMS08). Our model achieves excellent performance across multiple tasks while exhibiting superior multi-task coordination capabilities compared to existing approaches.