Explainable AI Enhances Trust in Heat Demand Forecasting Models

The increasing reliance on Machine Learning models for critical applications, such as heat demand forecasting in intelligent District Heating Systems, necessitates a strong focus on interpretability and trustworthiness. This paper proposes an ante-hoc Explainable AI (XAI) methodology to evaluate the global feature importance within these models. The goal is to enhance transparency, facilitate adherence to communal standards, improve customer satisfaction, and mitigate liability risks associated with opaque AI decisions. The methodology integrates four distinct approaches: the intrinsic interpretability offered by Gradient Boosting methods, alongside selected post-hoc techniques including Partial Dependence, Accumulated Local Effects (ALE), and SHAP. A key advantage of these chosen methods is that they do not rely on feature permutation or perturbations, which can introduce bias through the generation of unrealistic data values. The paper provides a detailed discussion of the results, highlighting the complementarities of these methods and offering specific interpretations tailored to the context of the district heating processes.