An interpretable machine learning approach for forecasting the occurrence of extreme electricity prices in the day-ahead market

In recent years, extreme electricity price events have become more frequent and severe due to increased renewable energy integration and external market shocks, posing significant risks to market participants and system operators. Accurate forecasting of such events is essential for supporting market stability, operational planning, and risk management. This paper proposes an adaptive dynamic weighted threshold (DWT) approach to identify extreme price events under dynamically changing market conditions. To address the challenge of imbalanced data, a weighted-XGBoost (W-XGB) classification model is developed to forecast extreme price occurrences. Comparative experiments demonstrate that the proposed method outperforms baseline models in predictive performance and class discrimination. To enhance interpretability, SHAP (SHapley Additive exPlanations) is applied to quantify feature contributions and provide insights into the drivers of extreme price events. Results show that extremely high prices are influenced by multiple factors, including supply-demand dynamics, fuel price volatility, and historical patterns, while extremely low prices are primarily driven by residual load, reflecting a more deterministic structure. The proposed framework integrates adaptive thresholding, imbalanced learning, and interpretable analysis, providing a practical tool to support decision-making in electricity markets. It offers actionable insights to inform bidding strategies, operational decisions, and risk management in increasingly volatile electricity markets.