In dynamic building environments, machine learning models used for energy management often suffer from performance degradation due to seasonality and evolving user behavior. This paper presents an adaptive retraining strategy that combines model performance evaluation with explainability-driven analysis to detect behavioral drift. By monitoring prediction accuracy and feature importance consistency, the system identifies when models require updating and classifies the urgency of retraining into three priority levels. Scheduling decisions are guided by contextual factors, such as energy consumption patterns and renewable energy availability, enabling timely retraining that supports continuous adaptation and aligns with green computing principles by minimizing both environmental and computational impacts. To validate the proposed strategy, a case study was conducted in a real smart office building environment. Experimental results show that the adaptive retraining strategy consistently outperforms a non-retrained baseline, reducing Weighted Absolute Percentage Error (WAPE), Symmetric Mean Absolute Percentage Error (SMAPE), and Root Mean Squared Percentage Error (RMSPE) by up to 5%. These improvements reflect enhanced predictive accuracy and robustness under shifting consumption patterns.