Accurate prediction of short-term power load plays an important role in grid planning, reducing power generation costs and improving power quality. In order to greatly improve the accuracy of short-term power load forecasting, an improved particle swarm optimization algorithm (IPSO) is used in this paper to optimize the long and short-term memory network (LSTM) to form a new power load forecasting model (IPSO-LSTM). LSTM is used to predict the power load, but it will encounter the problem of difficult selection of hyper-parameters. IPSO, which can effectively find the global optimal solution, is used to solve this problem. The inertia weight and learning factor in the particle swarm are fixed, which is easy to fall into the local optimum too early and miss the global optimum. The inertia weight and learning factor are changed from a fixed value to a non-linear law to balance its global search ability and local optimization ability. Through simulation analysis of actual case data, and comparing with PSO-LSTM, LSTM and Back Propagation (BP), the results verify that the prediction effect of the model in this paper is better. Experiments show that the proposed power load forecasting model has good accuracy and stability.