<![CDATA[Physics-based electricity price forecasting models, such as Linear Optimal Power Flow (LOPF), provide strong theoretical accuracy but often fail to capture real market prices. To reduce this gap, machine learning-based correction models are commonly applied; however, their reliability under significant market regime shifts remains unclear. This study investigates the robustness boundaries of price correction models using German electricity market data from 2019 and 2024, consisting of 17,544 hourly observations. Three key findings are identified. First, model selection and hyperparameter optimization do not improve cross-regime robustness: all 16 tested configurations fail when models trained in high-price regimes are applied to low-price regimes, increasing forecasting errors by 2.2–4.0 times. Second, the main source of failure is the target formulation. For the best-performing configuration, replacing the conventional price-difference target (EUR/MWh) with a ratio-based target turns all four transfer scenarios positive, achieving up to a 34.1% improvement in backward transfer. Third, model degradation is primarily driven by the distance between fuel-cost market regimes rather than temporal distance. Evaluation across 30 regime transitions shows that conventional correction approaches deteriorate as market condition differences increase. These findings reveal the limitations of current AI-based electricity price correction frameworks and highlight target redesign and regime-aware modeling as critical strategies for improving robustness under evolving market conditions.This study tests model resilience during the 2022 extreme market volatility (peak of the European gas crisis, RMSE: 159.06 EUR/MWh). The results show that the ratio target outperforms the difference target by +8% to +11% when predicting crisis periods. However, the model fails when trained on crisis data but tested under normal conditions, showing an explicit regime limit.]]>
Copyrights © 2026
