A hybrid data-driven methodology for forecasting normal and extreme electricity prices

Electricity price spikes are difficult to forecast but critical for market stability. These events arise from factors such as renewable energy variability, demand shocks, limited storage availability, market manipulation, and geopolitical disruptions. This paper presents and extends a hybrid electricity price forecasting methodology that integrates classification and forecasting. The work presented in this paper focuses specifically on the positive-price regime and further distinguishes between normal and extreme price days. To this end, Sparse Filtering is employed for feature extraction, while a Probabilistic Neural Network (PNN) is used for classification. Depending on the predicted category, electricity prices are subsequently forecast using an Extreme Learning Machine (ELM) combined with Bootstrap aggregation. The proposed approach was evaluated using data from the German and Finnish Day-Ahead electricity markets spanning 2020–2024 and demonstrated robust forecasting performance under normal, extreme, and negative price conditions, provided that the classification stage correctly identified the corresponding price regime. These results indicate that the proposed methodology provides a comprehensive and effective tool for managing electricity price volatility through forecasting.