An Optimized Ensemble Learning Framework for Energy Efficiency Assessment in Low-Voltage Distribution Networks Using Multi-Source Data Integration

This study proposes an optimized ensemble learning framework for energy-efficiency assessment in low-voltage distribution networks by integrating multiple data sources. The framework integrates heterogeneous data from smart meters, SCADA systems, meteorological stations, and network topology databases, employing advanced feature engineering to extract 89 essential predictors from 147 initial features. Three gradient boosting algorithms—Random Forest, XGBoost, and LightGBM—are combined through an elastic net stacking strategy with Bayesian hyperparameter optimization. The stacking ensemble achieved superior performance with an MAE of 118.4 kWh, an RMSE of 164.2 kWh, an MAPE of 3.98%, and an R2 of 0.952, representing 16.8% improvement over individual models. SHAP analysis provided model interpretability, identifying temperature, historical consumption, and temporal features as the primary drivers of efficiency. The framework demonstrated robust performance under data quality degradation and successfully generalized across diverse network configurations. Field implementation yielded an 8.3% reduction in distribution losses (95% CI: 7.2%–9.4%, p < 0.0001), 34% decrease in transformer failure rates (95% CI: 28%–40%, p = 0.003), and 12%–15% operational cost reduction. The framework’s ability to provide accurate predictions from 15 min to 24 h ahead while maintaining computational efficiency enables proactive distribution network management, supporting the transition toward efficient and sustainable power systems.