Graph Spatiotemporal Attention-Based Collaborative Demand Forecasting for Multi-Level Supply Chains

To address the coupling characteristics of multi-node demand in multi-level supply chains across network structure and temporal dynamics, this paper proposes a demand collaborative forecasting model based on a graph spatiotemporal attention mechanism. This model simultaneously characterizes structural dependencies, temporal correlations, and forecasting uncertainties within a unified framework. The method abstracts the supply chain as a weighted graph structure and performs dynamic relationship estimation driven by input states to form time-varying adjacency weights, thereby enhancing the ability to express the intensity and direction of cross-level transmission. In spatial modeling, hierarchical spatial attention is introduced to differentiate and aggregate information from different neighborhoods, highlighting the impact of key links and nodes on demand changes. In temporal modeling, temporal attention is used to select interpretable time segments within a historical window, and spatiotemporal coupling gating is employed to fuse structural signals and temporal representations to suppress noise interference and improve representation consistency. At the output layer, a multiquantile forecast head is constructed to generate different cross-sections of demand distribution, and quantile loss and monotonicity constraints are jointly optimized to avoid quantile overlap and improve interval reliability. Comparative experiments using an evaluation index system consistent with the probabilistic forecasting objective verify the effectiveness of the proposed method in characterizing supply chain demand distribution and improving interval reliability.