Scientific and accurate prediction of medium and long-term power demand is conducive to scientific planning of the power industry and coordinated development of the economy and society. By taking into account the 24 factors in six categories, such as macroeconomics and industrial structure, which affect the electricity consumption, the economic characteristics are fully reflected, and the multicollinearity among the factors is effectively eliminated by factor analysis. A relatively simple limit learning machine is set up based on parameters, and an improved particle swarm optimization model of limit learning machine is proposed. Through empirical analysis, the model is used to predict the historical load sequence to verify the effectiveness of the model.