Abstract: Short-term load forecasting (STLF) plays a crucial role in the operation and control of power systems. The stochastic nature and complexity of load pose challenges to accurate load forecasting. In this paper, the multi-kernel extreme learning machine (MKELM) optimized by a combination of fuzzy C-means clustering (FCM) theory, variational modal decomposition (VMD) and chaotic particle swarm optimization (CPSO), is introduced into the prediction model, to construct a load forecasting model of clustering, decomposition, optimisation, training, and prediction. Based on the application case of short-term load forecasting for the energy system of Tai Shan Station in the Chinese Antarctic interior, a load forecasting model was improved from the original model, which is applicable to domestic electricity consumption in China is obtained on the basis of the improvements in the original model. A comparison of the model training results shows that the new model has high accuracy in short-term load forecasting and can reflect the trend of regional electricity load, and the research results provide new methods and ideas for electricity load forecasting in various scenarios.