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

Area of Science:

• Electrical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Existing load forecasting methods struggle with multimodal data fusion and dynamic grid topology adaptation.
• Traditional methods often use rigid modality weights, leading to performance degradation.

Purpose of the Study:

• To propose a Dynamic Topology-Aware Multimodal Hypergraph Fusion Network (DTA-MHFN) for improved load forecasting in novel power systems.
• To address limitations in multimodal data fusion, dynamic topology adaptability, and modality weight allocation.

Main Methods:

• Constructed a Multimodal Hypergraph Network (MHN) modeling consumption and topology data as hypernodes with structural and behavioral hyperedges.
• Developed a Dynamic Topology Awareness Module (DTAM) using recurrent units to adaptively update hypergraph convolution weights based on topological changes.
• Implemented a Gradient-guided Adaptive Modality Weight (GAMW) mechanism for dynamic fusion weight allocation via attention.

Main Results:

• Achieved a 7.2% reduction in Mean Absolute Percentage Error (MAPE) on IEEE 33-bus and smart meter datasets.
• Improved prediction stability in dynamic topology scenarios by 12%.
• Demonstrated superior performance compared to existing mainstream load forecasting methods.

Conclusions:

• The DTA-MHFN effectively overcomes limitations of existing methods for load forecasting in novel power systems.
• The model provides significant theoretical and practical value for multimodal data fusion and adaptation to time-varying topologies.
• Offers technical support for reliable load forecasting in evolving power infrastructures.