Accurate prediction of short-term load provides basis for power dispatching of power plants and improves the economy of power systems. Due to the non-linear and non-stationary nature of load data, this paper proposes a prediction model based on EMD-IPSO-LSTM. First, the empirical mode decomposition (EMD) is used to deal with the nonlinear load sequence, and the sequence is decomposed into multiple intrinsic mode functions (IMF) and residuals (RES). The nonlinear decreasing assignment method and sine function are introduced to improve the inertia weight and learning factor of particle swarm optimization (PSO) respectively, so that the optimal solution of LSTM parameters can be found more effectively. Secondly, the IPSO is used to optimize the parameters such as the number of neurons in the first layer of LSTM, the loss rate, and the batch size. All IMF and res are divided into three groups: high, medium, and low frequency components, and are substituted into the optimized LSTM network for prediction. The final prediction result is obtained by superposition. Finally, a simulation experiment is carried out with the gefcom2014 power load forecasting competition data set, and compared with MLP, LSTM, PSO-LSTM, IPSO-LSTM, EMD-LSTM, EMD-PSO-LSTM. The results show that the proposed forecasting model has higher forecasting accuracy.