Load power fluctuation characteristics differ in time and space owing to many influencing factors. It is of great importance to determine the input features for load power modeling. This paper focuses on the feature selection of short-term power models. The purpose of this paper is to find the optimal set from the features including historical load, weather and date. Firstly, maximum information coefficient, recursive feature elimination method based on support vector machine and random forest are used for feature selection respectively. Secondly, the optimal feature set search strategy based on genetic algorithm is proposed according to the contrastive analysis. Finally, the optimal set of input features is finally determined. An example of calculation is carried out based on the bus load data of a 220 kV substation in a certain area. Compared with the load forecasting results of each feature selection method, the effectiveness and accuracy of the proposed feature selection method in short-term load forecasting have been verified.