Integrating PNN classification and ELM-Bootstrap for enhanced Day-Ahead negative price forecasting

Negative electricity prices occur under oversupply conditions, often driven by high renewable generation and the limited flexibility of conventional power units. Without sufficient energy storage, surplus electricity must be absorbed by the grid, leading to price drops below zero. Accurate forecasting of such events is critical, yet current literature offers limited solutions focused specifically on negative price prediction. This paper proposes a novel two-stage hybrid methodology for forecasting negative Day-Ahead electricity prices. In the first stage, an Extreme Learning Machine combined with the Bootstrap method generates prediction intervals. In the second stage, a Probabilistic Neural Network classifies the target day as either a negative or positive price day, incorporating market-based misclassification costs to reflect the differing economic impacts of price directions. The final forecast selection depends on the PNN outcome: if a negative price is predicted, the lower bound of the ELM-Bootstrap interval is used; if positive, the average is taken. The PNN classifier is benchmarked against Support Vector Machine with an RBF kernel, Extreme Gradient Boosting, and Long Short-Term Memory models, demonstrating superior classification accuracy. The proposed hybrid approach also outperforms the standalone ELM-Bootstrap method in forecasting precision. The methodology is applied to the German Day-Ahead market using data from 2020 to 2023, a period marked by high market volatility, validating its effectiveness under real-world conditions.