Transformer-GNN fusion with gradient surgery for electricity price forecasting

Area of Science:

• Electrical Engineering
• Data Science
• Computational Economics

Background:

• Accurate short-term electricity price forecasting is vital for market efficiency and grid stability.
• Existing methods often fail to capture complex temporal and spatial dependencies inherent in power systems.
• Modeling network constraints and long-range temporal patterns remains a significant challenge.

Purpose of the Study:

• To develop a hybrid forecasting framework that effectively models both temporal and spatial dynamics in electricity markets.
• To integrate market signals with auxiliary physical data for enhanced forecasting accuracy.
• To validate a novel approach for cross-domain spatiotemporal fusion using simulated physical data.

Main Methods:

• A hybrid model integrating a Transformer encoder for temporal dependencies and a Graph Neural Network (GNN) for spatial correlations.
• Utilizing PCGrad gradient surgery to manage training conflicts between heterogeneous model components.
• Combining PJM market data with simulated IEEE 39-bus PMU data via synchronized alignment and attention-based fusion.

Main Results:

• The proposed model demonstrated superior performance compared to classical and hybrid benchmarks on fused datasets.
• Achieved the lowest average error across five independent runs among fused-data methods.
• Exhibited narrower uncertainty intervals than the leading multimodal baseline, indicating improved prediction reliability.

Conclusions:

• Temporally aligned fusion of market signals and auxiliary physical features enhances electricity price forecasting accuracy.
• The hybrid framework offers a robust method for integrating diverse data sources in power system modeling.
• This proof-of-concept highlights the potential of physically informed auxiliary data for improving forecasting in complex energy systems.