Interpretable Smart Meter Anomaly Detection Based on Bayesian-Optimized XGBoost and SHAP

Smart meter anomaly detection is critical for ensuring the security and stability of smart grids. However, existing detection methods face the following limitations: insufficient feature extraction, severe class imbalance, inefficient manual hyperparameter tuning, and poor model interpretability. To overcome these limitations, we propose an accurate and interpretable anomaly detection method based on Bayesian-optimized XGBoost and SHAP. Our method integrates multi-dimensional feature extraction to enrich feature information, Borderline-SMOTE to mitigate class imbalance, Bayesian optimization to tune XGBoost hyperparameters, and SHAP to quantify feature contributions and provide model interpretability. Experimental results on the public MAD dataset demonstrate that our method consistently outperforms both classical machine learning models, including decision tree, Random Forest, XGBoost, and LightGBM, as well as representative deep learning models such as CNN, TCN, LSTM, and CNN-LSTM in binary and multi-class classification tasks, achieving superior accuracy, precision, recall, and F1 scores. SHAP analysis further reveals that three-phase unbalance features are the dominant indicators of abnormal samples, a finding highly consistent with the physical mechanisms of power systems. Our method achieves both competitive detection performance and transparent decision-making, providing an actionable solution for smart meter anomaly detection in practical engineering applications.