In order to improve the efficiency of distribution network losses analysis and evaluation on a distribution network, this paper proposes a distribution network loss prediction method based on feature selection and Stacking integrated learning. First, based on the main method of feature selection, through the correlation analysis method, the maximum information coefficient method and the tree-based model, the input features are comprehensively analyzed and the importance of various features for the network losses prediction is obtained. The important features are selected as the input of the distribution network losses prediction model. We then introduce Stacking integrated learning theory, which considers the advantages of combining multiple prediction models. We establish the Stacking integrated learning distribution network losses prediction model and finally obtain the results through simulation. The simulation data of the paper is derived from the real data of 44 stations in a line of a 10 kV distribution network in Hunan Province. The results show that the method can effectively improve the accuracy and robustness of distribution network losses prediction. Compared with a single prediction model, it has higher prediction accuracy and better generalization capabilities.This work is supported by National Natural Science Foundation of China (No. 51822702), and Science and Technology Project of State Grid Hunan Electric Power Company (No. 5216A517000U), and the Excellent Innovation Youth Program of Changsha of China (No. KQ1802029).