Prediction accuracy of short-term load is critical to the normal operation of the microgrid due to the strong randomness of load. A kernel extreme learning machine (KELM) prediction model based on complementary ensemble empirical mode decomposition (CEEMD) and regional-division self-adapting variation particle swarm optimization (RSVPSO) is proposed. The load sequence is decomposed into several smooth subsequences by using complementary ensemble empirical mode decomposition to reduce the mutual influences among different local information. Aiming at the problem that particle swarm optimization is easy to fall into local optimization and is slow in converge, a inertial weight and learning factor based on regional-division are utilized to improve the global search ability and search efficiency, further, adaptive variation operation is introduced to avoid the population falling into local optimum. The prediction accuracy of kernel extreme learning machine is obviously improved. Finally, the model proposed in this paper can obtain good performance of accuracy about 98.114%, which has better prediction effect and practical application significance than other prediction models.