Load forecasting is an important aspect of ensuring power supply-demand balance in distribution station and has significant guiding significance for the safety warning and stable operation of the power system. Affected by various other factors, the direct prediction model of the load in the distribution area usually has poor generalization ability and is difficult to meet the load forecasting requirements of complex distribution substations. To improve the generalization ability of load forecasting methods in distribution areas, a load prediction method based on multi-source environmental feature fusion is proposed. Firstly, the external environmental features are analyzed by principal component analysis, and the input variables are reduced and corrected to extract the environmental feature components that affect the load changes in distribution area. Based on the proposed load identification model, the substation load is identified by feature fusion. Then, based on the comprehensive analysis of the identified substation load and environmental feature components, the proposed load forecasting model is used to predict the substation's individual load by feature fusion. Finally, the linear superposition of the predicted results of individual loads is obtained, and the load prediction result of the distribution substation is obtained. Selecting the load data of a low-voltage distribution substation in China for two years as an example, compared with other direct prediction methods, the proposed method's load prediction curve is closer to the true load data curve, effectively improving the accuracy of load forecasting.