Energy loss prediction using least absolute shrinkage and selection operator regression and SHAP explainability

Abstract

Technical energy loss estimation in power distribution systems is essential for improving operational efficiency and cost-effectiveness. However, distribution-level datasets are often failed to cope with the nonlinear behavior and sparse, low-resolution data typical in modern grid environments. This study proposes an interpretable artificial intelligence (AI) framework based on least absolute shrinkage and selection operator (LASSO) regression integrated with Shapley additive explanations (SHAP) to estimate and explain technical energy losses at the feeder level. An exploratory multicollinearity assessment using correlation analysis and variance inflation factor (VIF) revealed severe redundancy among operational variables, justifying the adoption of L1-regularized regression. Hyperparameter tuning via LassoCV identified an optimal regularization parameter, resulting in strong predictive performance. Comparative evaluation with nonlinear models, including random forest and gradient boosting, demonstrated that LASSO achieves competitive or superior generalization performance while preserving interpretability. Feature importance analysis and SHAP-based explanations confirmed that operational loading variables particularly infeed energy, load factor, and maximum demand are the dominant drivers of technical losses. SHAP dependence and interaction analyses further revealed context-dependent behavior among correlated predictors, enriching interpretability beyond coefficient-based rankings. The results demonstrate that regularized linear modeling, when combined with explainable AI techniques, provides a robust, transparent, and practically deployable solution for technical loss estimation in distribution networks.