The invention discloses a short time electric power load prediction method and device. The short time electric power load prediction method predicts an electric power load through integrating quantile regression and robustness extreme learning and using a hybrid prediction model established after optimizing a hybrid particle swamp algorithm (PSOGSA). The quantile regression uses multiple quantiles of history electric power data influence factors to obtain a quantile equation corresponding to a conditional distribution of electric power load prediction data of sometime in the future; stochastic disturbance of electric power data inputted in quantile regression can describe statistic distribution of prediction load values in detail without the need of making any hypothesis on distribution, which makes the whole prediction model strong in robustness; the robustness extreme learning machine is more robust in an abnormal load value; and the electric power load can be predicted by combining the two methods and using a hybrid model formed after optimizing the PSOGSA. The invention also provides a short time electric power load prediction device having the same beneficial effects.