Power load exhibits characteristics of temporal and spatial variation and is affected by various factors. In short-term load forecasting, an excessive number of input features can cause dimensionality disasters and lead to poor model prediction performance. Therefore, selecting a reasonable input feature set is crucial. This article proposes a novel feature selection method for short-term load forecasting–the mRMR-IPSO two-stage method. The max-relevance and min-redundancy (mRMR) criterion is employed to rank the original features, considering both the correlation between input and output features and the redundancy among input features. This process filters out less impactful features ranked lower and initially selects these significantly influencing the prediction. Then, an improved particle swarm optimization (IPSO) algorithm-based search strategy is adopted. The prediction accuracy of the LightGBM model is used as the fitness function during the search, facilitating the selection of primary feature subsets and obtaining optimal feature subsets. Calculation examples show that the proposed method improves prediction accuracy while substantially reducing the original feature set.