Accurate residential energy forecasting faces critical challenges regarding privacy, evolving load patterns, and model transparency. We present EXFEDL, an explainable clustered federated learning framework that groups households by temporal similarity and learns cluster-specific models using a temporal convolutional network with global attention. To enhance interpretability, Shapley Additive explanations provide insights at both household and cluster levels. A stability controller monitors behavioral drift and triggers adaptive re-clustering to maintain reliability over time. Experiments on real-world datasets show that EXFEDL improves forecasting accuracy, explanation consistency, and adaptability compared to centralized and standard federated learning approaches.
