A Spatiotemporal Coupled Prediction Method for Multi-Energy Flows in Transportation Hubs Based on H3 Grid Representation and Heterogeneous Graph Attention–Transformer Networks

Large transportation hubs such as ports are becoming high-density agglomerations of energy consumption, in which electricity, hydrogen, and thermal/cooling energy flows are increasingly coupled with traffic activity, so their coordinated and low-carbon operation has become an important concern for both intelligent transportation and integrated energy systems. However, the multi-source data of such hubs differ greatly in spatial granularity and physical meaning, and traffic and energy loads are jointly coupled across space and time, which makes conventional methods that forecast a single system unable to characterize the overall dynamics. To address this problem, this paper proposes a heterogeneous data fusion and spatiotemporal coupled prediction method for multi-energy flows in transportation hubs. A unified representation framework based on H3 hexagonal grids is first designed, which aligns multi-source data onto a common spatial index and fuses theminto a standardized fourth-order tensor through area, length, and frequency spatial weights together with quantile truncation. On this basis, a hybrid model that couples a heterogeneous graph attention network (HetGAT) with a Transformer encoder is constructed, in which HetGAT models the heterogeneous spatial topology formed by four node types and five edge types, while the Transformer encoder extracts long-range temporal dependencies. An uncertainty-weighted dual-task head then jointly outputs traffic states and multi-energy loads. Experiments on a Ningbo-Zhoushan Port scenario built with the SUMO simulation platform show that the proposed method achieves R2 values of 0.9763, 0.8088, 0.9644, and 0.9583 for traffic flow, electricity load, hydrogen consumption, and thermal/cooling load, respectively, while consistently outperforming single-system and homogeneous-graph baselines on the energy-side prediction tasks.