The importance of precise long-term forecasting in practical applications continues to rise. Extensive scenarios, including parking resource prediction and environmental quality monitoring, rely significantly on LSTF's accurate spatio-temporal forecasting capabilities. This technology strengthens prediction effectiveness by combining interaction relationships between spatial-temporal dimensions with contextual data integration. Over time, graph neural networks (GNNs) have proven highly effective in capturing spatial interdependencies. Recent advances have introduced multi-GNNs (MGNNs), which incorporate more contextual insights to improve predictive accuracy. However, when MGNNs are applied to long-term spatio-temporal forecasting (LSTF), they encounter challenges such as limited generality, under-utilization of context, static graph merging methods, and overlooking dynamic interrelations. To address these issues, we propose novel graph structures that encode each node's contextual information while fully exploiting long-term spatio-temporal dependencies. Furthermore, this research designs a dynamic multigraph fusion architecture that integrates spatial dimensions, temporal features, and graph attention mechanisms to simultaneously capture intragraph node correlations and cross-graph interactions. To strengthen relational analysis, trainable weight tensors are employed for quantitative evaluation of node importance across graphs. Systematic experiments on three large-scale benchmark datasets confirm that this approach achieves significant performance enhancement for existing GNNs in LSTF tasks.
Xiao et al. (Thu,) studied this question.