Abstract: To address electricity undercounting caused by metering device faults, this paper proposes a counterfactual electricity reconciliation method based on line loss analysis and ensemble learning. The method solves the shortcomings of traditional reconciliation methods, including low fault location accuracy, large prediction errors of theoretical line loss, and poor time-of-use splitting performance for long-period faults. It constructs a complete technical framework of accurate fault location, dynamic line loss rate prediction, daily recovered electric energy calculation, and intelligent time-of-use electric energy splitting. A sliding-window Z-score method is used to automatically detect the start and end times of faults. An XGBoost model is used to predict the dynamic theoretical line loss rate, and SHAP values are introduced to enhance interpretability. Daily recovered electric energy is calculated based on counterfactual logic, and an Informer model is introduced to achieve accurate time-of-use splitting of electric energy for long-period faults. Case verification shows that the MAPE of XGBoost line loss prediction is 4.2%, and the accuracy is 51.7% higher than that of a linear model. The time-of-use splitting accuracy of Informer is 18.3% higher than that of the similar day method. The proposed method provides technical support for electricity reconciliation under metering faults.