Endogenous regime switching in day-ahead electricity price forecasting: An economic physics-guided residual regime network approach
Applied Soft Computing, vol.203, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 203
- Publication Date: 2026
- Doi Number: 10.1016/j.asoc.2026.115931
- Journal Name: Applied Soft Computing
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Compendex, INSPEC
- Keywords: Deep Learning, Electricity Price Forecasting, Physics-Informed Neural Networks, Regime Switching, Renewable Integration
- Erciyes University Affiliated: Yes
Abstract
Day-ahead electricity price forecasting is increasingly challenged by the structural volatility introduced by renewable energy sources. Existing literature largely relies on a bifurcation strategy, treating price spikes as statistical outliers to be filtered or handled by separate ensemble modules. We argue this approach creates a bifurcation fallacy, severing the causal link between the physics of the economy and price formation. To address this, we propose the Physics-Guided Residual Regime Network (PGRRN). Unlike static architectures, PGRRN utilizes a soft regime-switching mechanism that decomposes the prediction pathway into a linear base-load component and a non-linear congestion component. A physics-guided gating unit, conditioned on exogenous load and temporal forecasts, dynamically weights these components, allowing the model to endogenously adapt to market regimes. Experimental results on five major market datasets (Belgium, Germany, France, Nord Pool, and PJM) demonstrate that PGRRN significantly outperforms robust XGBoost, DNN, and LEAR models. For instance, in the Nord Pool market, PGRRN achieves a Mean Absolute Error (MAE) of 0.92 compared to 1.68 for the best of the DNN ensembles. While the method incurs a higher computational cost than LEAR ensembles, it is significantly lower than that of DNN ensembles. The significant improvement in capturing structural volatility confirms the efficacy of embedding economic physics into deep learning architectures.