Integrating bidding data into electricity price forecasting: An interpretable graph representation modeling approach

Bidding data plays a critical role in electricity price formation. However, existing electricity price forecasting (EPF) research often fails to incorporate it directly due to its inherent characteristics, such as heterogeneity, nonlinearity, high dimensionality, and redundancy. To address this issue, we propose an interpretable graph representation modeling approach, named Graph-X, to effectively integrate bidding data into EPF. Individual-level bidding data are first aggregated into system-level curves and then transformed into bidding graphs based on the marginal clearing principle, ensuring that each graph element retains a clear market interpretation. On this basis, a tailored neural architecture is developed to extract price-relevant features from bidding graphs and directly map them to price forecasts. The model combines sparse convolution for structurally efficient spatial feature extraction with a lightweight GRU for temporal dynamics, enabling both accurate forecasting and transparent feature attribution. Evaluated on ISO New England data, Graph-X outperforms benchmarks, reducing forecasting errors by up to 16.91% in terms of MSE. The results demonstrate that graph-based preprocessing and the tailored neural architecture enhance accuracy by focusing on supply-demand intersections, offering a robust end-to-end solution for EPF.