AI-Driven Probabilistic Day-Ahead Electricity Price Forecasting with Tail Risk Mitigation via Dynamic Spike Calibration

Accurate day-ahead electricity price forecasting remains challenging due to tail risks induced by market volatility. Conventional forecasting models often fail to capture price spikes associated with extreme price values, leading to significant tail bias. This paper presents a robust probabilistic price forecasting framework that integrates extreme value theory with deep-learning probability forecasting and post-hoc calibration. Firstly, a quantile-based sliding window is developed to identify extreme price realizations, and generalized extreme value distribution parameters are estimated using a gradient-based optimization approach for extreme value fitting. Meanwhile, a robust Box-Cox transformation and Huber loss are jointly employed to improve distributional modeling and training stability. Finally, a dynamic calibration mechanism is introduced to refine extreme price predictions and enhance the quantification of tail uncertainty. Case studies using real-world electricity market data from Guangxi, China, demonstrate the effectiveness of the proposed framework. Compared to state-of-the-art baselines, the proposed framework improves extreme-value prediction accuracy with an R2 increase of 0.46, and the average coverage errors at the 80%, 90%, and 95% confidence levels are reduced to 6.00%, 3.57%, and 2.98%, respectively. Additional experiments using Transformer and temporal convolutional network backbones further verify the strong generalization capability. © 1975-2011 IEEE.