Enhancing explainability in AI-based energy forecasting through clustering and data selection

Explainable Artificial Intelligence (XAI) seeks to enhance the interpretability of Artificial Intelligence (AI) systems, ensuring that algorithmic decisions and their underlying data are comprehensible to non-technical stakeholders. While advanced Machine Learning (ML) models, such as deep neural networks, have significantly improved AI capabilities, their complexity poses challenges for XAI, particularly in handling large datasets required for training and interpretation. In particular, the application of Shapley Additive Explanations (SHAP), although widely recognized for its effectiveness, often incurs a high computational cost when applied to large-scale data. Addressing this issue, our previous work proposed a novel approach that leverages K-Means clustering to identify representative data instances, applied after the forecasting phase to refine SHAP-based explanations and reduce computational costs while preserving their fidelity. This extended study further optimizes the clustering strategy and evaluates its applicability across broader use cases in sustainable energy systems. We apply our method to forecast photovoltaic (PV) generation in buildings, a critical aspect for energy management in e-mobility and smart grids. The results show that clustering reduces execution time by more than 50 % compared to random sampling while maintaining comparable explanatory stability. These findings highlight the potential of data-driven clustering techniques in enhancing the explainability of ML models in energy forecasting, contributing to more accessible and practical AI solutions for real-world applications. • K-Means-based approach proposed to optimize explainability in XAI models. • Considerable data reduction without compromising the quality of explanations in complex machine learning solutions. • Application in forecasting photovoltaic energy generation for sustainable management of electric mobility. • Results that support the solution’s ability to reduce execution time while maintaining accuracy significantly. • Improving AI interpretability in real-world smart energy scenarios.