A Mode-Decomposition-Based Semantic Fusion Network with Graph-Guided Spatial Attention for Urban Load Demand Forecasting

Abstract

Accurate urban load demand forecasting is critical for the economic dispatch and reliable operation of power systems. However, urban load forecasting remains challenging because forecasting models must extract reliable representations from entangled multi-scale dynamics, temporal semantic dependencies, and complex spatiotemporal correlations. To address these issues, this paper proposes a mode-decomposition-based semantic fusion network with graph-guided spatial attention (MDSFGA) for urban load demand forecasting. Specifically, a multi-scale spatiotemporal encoder is developed to decompose historical observations into globally frequency-consistent mode components across spatial nodes and to construct corresponding spatiotemporal representations through branch-wise feature extraction and an adaptive mode feature fusion module. Within this encoder, a semantic-aware temporal extractor incorporates temporal semantic context into recurrent dependency modeling. At the same time, a graph-guided spatial attention module combines mode-specific graph priors with spatial self-attention to jointly model local structural constraints and global spatial dependencies. Experiments on four public load datasets demonstrate that MDSFGA outperforms representative baselines. Notably, MDSFGA achieves a MAPE of 15.65% on the CELP-VA dataset, with a 7.83% improvement compared with the second-best baseline.