Predicting faults in power distribution grids during heatwaves: A comparative study of machine learning models applied to Milan distribution network

Abstract

The increasing frequency and severity of extreme weather events, such as heatwaves in Milan, intensified by climate change, pose significant challenges to the reliability and resilience of electrical power distribution systems. Traditional deterministic planning methods are becoming inadequate as these events grow more unpredictable. This study introduces a novel machine learning methodology to enhance grid resilience during heatwaves, focusing on fault prediction and heatwave forecasting. Three complementary approaches were systematically evaluated: Ridge Regression with Recursive Feature Elimination and Cross-Validation, Random Forest Regression, and Second-order Polynomial Poisson Regression with Recursive Feature Elimination and Cross-Validation. Through innovative feature engineering incorporating soil temperature, humidity gradients, and dynamic load demand patterns, predictive accuracy was significantly improved over conventional methods. Rigorous cross-validation with statistical validation demonstrated model stability across varying conditions, with the Second-order Polynomial Poisson model achieving a mean absolute error of 0.15 in predicting fault occurrences. To address the observed heteroscedasticity during high-fault periods, adaptive prediction intervals were developed, providing operators with crucial uncertainty quantification when it matters most. When translated to operational reality, these models enable Distribution System Operators to implement proactive fault management strategies, potentially reducing outage response times by an estimated 15–20 % during extreme weather events. This research bridges the critical gap between climate science and power system engineering, offering data-driven decision support for the increasingly volatile operational environment facing urban distribution networks.