The embodiment of the invention provides an EMD-GRU short-term power load prediction method based ON feature selection. The comprises the following steps: decomposing an original load sequence into aplurality of time sequence components by using an EMD (empirical mode decomposition) method, taking all the time sequence components as an initial feature set, and forming a potential input variable of a prediction model by using the initial feature set; carrying out correlation analysis on the initial characteristics through a Pearson correlation coefficient method, and selecting a time sequencecomponent with relatively high correlation with the original load sequence as an input characteristic of a prediction model; and inputting the selected time sequence components in combination with theoriginal load sequence into a GRU prediction model, and executing final load prediction. According to the technical scheme provided by the embodiment of the invention, the accuracy of short-term power load prediction can be improved.