Ke Zhang ^* Yongwang Zhang Jian Li ^* Zetao Jiang ^* Yuxin Lu Binghui Zhao ^*

• State Grid Corporation of China (SGCC), Beijing, China

The line losses in distribution networks greatly affect the quality of grid operation, and accurate prediction of these losses can effectively facilitate power system planning and network restructuring. Therefore, this paper proposes a distribution network line loss prediction method based on feature selection and Stacking ensemble learning to improve the effectiveness of distribution network loss analysis and assessment. Utilizing data from 44 substations over 18 months, we integrated a Stacking ensemble learning model with advanced feature selection techniques, including correlation coefficient, maximum information coefficient, and tree-based methods, to identify key predictors of power loss. The model achieved a Mean Absolute Percentage Error (MAPE) of 3.78% and a Root Mean Square Error (RMSE) of 1.53, indicating a significant improvement over traditional linear regression-based prediction method. The analysis highlights the importance of historical line loss and line active power as predictive variables, alongside the inclusion of time-related features for model refinement. The study demonstrates the efficacy of combining multiple feature selection methods with Stacking ensemble learning for predicting power loss in 10 kV distribution networks. The model enhanced accuracy and reliability offer valuable insights for electrical engineering applications, potentially contributing to more efficient and sustainable energy distribution systems.