Power load has the characteristics of time and space change and it is affected by many factors. In short-term load forecasting, excessive input features will cause dimensional disasters and lead to poor model prediction performance. Therefore, it is very important to choose a reasonable input feature set. The article proposes a new feature selection method for short-term load forecasting—mRMR-IPSO two-stage method. Use the Max-Relevance and Min-Redundancy (mRMR) criterion to sort the original features, consider the correlation between the input and output features and the redundancy between the input features, filter out some of the later ranked features, and initially select the ones that have a significant impact on the prediction effect Feature subsets; Then, a search strategy based on the improved particle swarm optimization (IPSO) algorithm is adopted, and the prediction accuracy of the LightGBM model is used as the fitness function to select the primary feature subsets to obtain the optimal feature subsets. The results of calculation examples show that the proposed method can improve the prediction accuracy while greatly reducing the original feature set.