The present invention relates to a short-term power load prediction method. The method comprises the following steps: (1), collecting circuit load data of a plurality of days and a plurality of time points in each day, to constitute an original load sequence; (2), performing collective empirical mode decomposition on the original load sequence, and adaptively decomposing the original load sequence into a series of intrinsic mode functions with specific periodic components and a residual amount; (3), using an extreme learning machine to perform prediction on each component obtained through decomposition, and using the extreme learning machine to perform prediction model modeling on load data corresponding to consecutive days and the same time points in each intrinsic mode component and residual component; and (4), performing power load prediction by using the obtained prediction model. According to the short-term power load prediction method, the problem of an end effect of traditional empirical mode decomposition is overcome. On the other hand, a generalization ability and a calculation speed of the extreme learning machine are better than those of an artificial neural network method that is commonly used at present.