Modern AI-enabled infrastructure systems increasingly rely on continuous sensing and data streams, yet most high-performing anomaly detection methods remain opaque and difficult to trust. This paper presents an explainable anomaly detection framework for large-scale smart-meter data that unifies the predictive strength of ensemble models with the transparency of interpretable machine learning. We first train a high-accuracy XGBoost/LightGBM ensemble on year-long hourly electricity consumption data from 200 buildings to produce reliable pseudo-labels for unlabeled meter readings. These labels are then used to train an Explainable Boosting Machine (EBM), a transparent, glass-box model that captures nonlinear consumption patterns while providing both local and global interpretability. The resulting two-stage pipeline delivers ensemble-level performance with full interpretability, enabling domain experts to visualize feature contributions, temporal patterns, and root causes of detected anomalies. Evaluated on the real-world multiple smart meter dataset, our system achieves high detection accuracy while offering human-understandable explanations at both model and instance levels. We further integrate the model into a real-time Streamlit dashboard for interactive anomaly exploration and sustainability monitoring, including energy and CO2 footprint tracking. This work demonstrates a practical engineering approach for bridging accuracy and interpretability in AI-driven energy analytics, advancing trustworthy deployment of explainable AI in critical infrastructure.