Explainable AI-Based Early Warning System for Electricity Price Spikes: Evidence from the Romanian Market

Electricity price volatility represents a critical challenge for market participants, grid operators, and regulators in liberalized energy markets. In the context of the European energy transition, understanding what drives sudden price spikes is essential for risk management and market stability. While machine learning models have shown strong predictive capabilities for electricity prices, most remain opaque, offering little insight into the underlying causes of their predictions. At the same time, explainable artificial intelligence methods have gained prominence across domains but remain underexplored in the specific context of electricity price spike detection and early warning. This study investigates whether explainable artificial intelligence can identify load-generation balance indicators as primary drivers of electricity price spikes in the Romanian day-ahead market. Using hourly market data covering March to September 2024 (5,000 observations and 70 engineered features), we train gradient boosting and ensemble tree models under a strict chronological validation framework that prevents data leakage. We then apply two complementary explanation techniques—one that reveals which factors matter most across all predictions, and another that explains individual price spike cases—to validate the role of supply-demand balance signals. Our results show that the best model explains 67.6 percent of price variance, with the residual load forecast emerging as the second most important predictor globally and the most important trigger in the early warning system. The composite warning system detects 72.2 percent of high-price events with an overall quality score of 0.57, which represents a strong result for a threshold-based early warning system in a setting where standardized benchmarks are scarce. As supplementary evidence rather than a separate research question, cross-validation shows that performance degrades during the summer 2024 heat wave, yet the balance indicators still signaled the stress conditions, highlighting the value of interpretable signals even when predictive accuracy declines. The main contribution of this paper is the development and validation of an integrated explainable artificial intelligence framework that combines global and local interpretability for electricity price spike early warning, providing market operators with transparent, actionable signals rather than black-box predictions.