•  
  •  
 

Keywords

calculation of distribution network line loss, feature selection, correlation analysis, LSTM-XGBoost-SHAP m odel

Abstract

Accurately calculating distribution network line losses is crucial for ensuring the economic，safe and stable operation of the power system.The conventional calculation methods cannot effectively utilize electrical feature parameters and involve cumbersome calculation procedures.To address this issue，the study proposes an interpretable calculation method for distribution network line loss based on electrical feature selection and a hybrid model integrating long short-term memory （LSTM ） and extreme gradient boosting-SHapley Additive exPlanations （XGBoost-SHAP ）.In this method，theoretical electrical feature indicators for line loss are determined by comprehensively considering the relationships between various electrical parameters and distribution network line losses.A feature selection method based on cross-correlation is proposed to eliminate redundant features and construct the optimal feature subset，thereby overcoming high redundancy among features of the same type.Then，an LSTM-XGBoost-SHAP interpretable calculation framework is developed to solve the problem of insufficient utilization of heterogeneous data by conventional single models.In this framework，the de ep features of time series data are extracted using the LSTM network，and the time series features are fused with static features by the XGBoost model to achieve high-precision calculation of the theoretical line loss.The contribution of each input electrical feature parameter to the theoretical line loss is analyzed using the SHAP module.Finally，the accuracy of the proposed method is verified using actual operational data from a distribution network in Northwest China.The results show that the proposed method achieves a root mean square error （RMSE ） of less than 0.213% and a mean absolute percentage error （MAPE ） of not more than 5.634%，indicating significantly better calculation accuracy than that of the conventional single models as well as favorable robustness and engineering applicability.

DOI

10.19781/j.issn.1673-9140.2026.04.007

First Page

75

Last Page

87

Download

Share

COinS