A Hybrid Classification-Regression Method for Forecasting Negative Electricity Prices

The rapid growth of solar generation in South Australia has led to more frequent occurrences of negative wholesale electricity price events, creating challenges for grid stability and reliability. By the fourth quarter of 2024, AEMO reported negative prices in 38% of South Australia's dispatch intervals. Accurate forecasting of these events is non-trivial due to the skewed price distribution, complex system interactions, and significant class imbalance. To address this challenge, we developed a two-stage hybrid method in which a classifier first estimates the probability of a negative price event, followed by a regressor that predicts the event's magnitude. Using a full year of AEMO 5-minute data, our results show that the proposed model outperforms comparable studies by capturing 98% of negative-price events. Experimental analysis further demonstrates that the method effectively learns temporal patterns and system dynamics while maintaining interpretability through SHAP and feature-importance analysis. An economic evaluation showed that the approach achieves 95.9% of the theoretical optimum, indicating near-optimal storage operation performance.