Detecting Concept Drift with SHapley Additive ExPlanations for Intelligent Model Retraining in Energy Generation Forecasting

Abstract

Accurate energy generation forecasting is essential for effective energy management. However, it remains a complex task due to the influence of dynamic factors such as meteorological conditions, seasonal variations, and evolving grid operations. Ensuring model reliability over time requires continuous assessment to detect performance degradation. Traditional retraining strategies, including periodic updates and statistical drift detection techniques, often struggle to balance model accuracy with computational efficiency. This study introduces a novel approach that leverages SHapley Additive Explanations (SHAP) to dynamically detect concept drift by analyzing variations in feature importance. The methodology establishes a baseline SHAP distribution and identifies deviations that indicate drift, prompting model retraining when necessary. A comparative evaluation is conducted against conventional methods, including scheduled retraining, Adaptive Windowing (ADWIN), and the Kolmogorov-Smirnov (KS) test. Furthermore, a sensitivity analysis examines the impact of key configuration parameters on detection accuracy and computational cost. The results demonstrate that SHAP-based drift detection improves forecasting accuracy, achieving a 26.67% to 35.29% reduction in Mean Squared Error, while maintaining an adaptive retraining strategy. These findings underscore the potential of SHAP as an interpretable and efficient approach for managing concept drift in energy forecasting applications.