Dynamic topology-aware multimodal hypergraph fusion network for load forecasting in novel power systems

Existing load forecasting methods for novel power systems face critical bottlenecks regarding insufficient multimodal data fusion along with weak dynamic topology adaptability and rigid modality weight allocation. A Dynamic Topology-Aware Multimodal Hypergraph Fusion Network (DTA-MHFN) is proposed to address these challenges. First, a Multimodal Hypergraph Network (MHN) is constructed. This network models user electricity consumption time series units and grid topology physical units as distinct hypernodes. Structural and behavioral hyperedges are then built based on physical connections and behavioral correlations. This design achieves explicit representation of cross-modal high-order associations and overcomes the limitation of traditional graph neural networks that only model pairwise correlations. Second, a Dynamic Topology Awareness Module (DTAM) is designed. It monitors topological time variations by calculating the cosine similarity of topological adjacency matrices across adjacent time windows. A gated recurrent unit is integrated to capture evolutionary features and adaptively update the hypergraph convolution weights. This mechanism resolves the issue of sudden prediction accuracy degradation under perturbation scenarios caused by static topology modeling. Finally, a Gradient-guided Adaptive Modality Weight mechanism (GAMW) is proposed. It quantifies the contribution of bimodal features using the proportion of backpropagation gradient magnitudes. Fusion weights are dynamically allocated through an attention mechanism. This strategy avoids feature redundancy or critical information loss triggered by static weights. Experimental results demonstrate that the proposed model reduces the MAPE by 7.2% on the IEEE 33-bus system and smart meter datasets. The prediction stability in dynamic topology scenarios improves by 12%. The comprehensive performance significantly outperforms existing mainstream methods. This provides technical support for load forecasting in new power systems and holds significant theoretical and practical value for advancing multimodal data fusion and adaptation to time-varying topologies.