To help overcome the shortcomings of short-term power load forecasting caused by random initialization of artificial neural network parameters, a forecasting method based on an improved multivariate universe optimizer (IMVO) algorithm and an extreme learning machine (ELM) is proposed. The improvement of the algorithm includes the following three aspects. First, the improved Tent chaotic mapping method is obtained by adding the random number of beta distribution, and the improved Tent chaotic mapping method with better ergodic uniformity is used to make the MVO algorithm obtain a good initial solution. Second, the travel distance rate of the traditional MVO algorithm is improved using the exponential form, and the improved algorithm can maintain a high global development level in the whole optimization iteration before and during the middle period. Then, the elite reverse learning method is used to improve the universe group. The performance of the algorithm before and after improvement is tested by the benchmark function, indicating that the IMVO algorithm has better stability and robustness. Finally, the IMVO algorithm is used to optimize the weights and thresholds of an ELM, and the IMVO-ELM short-term power load forecasting model is established. Case analysis and comparative experiments show that the stability, prediction accuracy and generalization ability of IMVO-ELM model are better than those of other models. This work is supported by the National Natural Science Foundation of China (No. 52177085).