Promotion and use of new energy power generation has aggravated the contradiction between supply and demand of power grid during peak hours. Identification of load patterns of power users can provide support for load participation in peak regulation decisions. In order to improve the accuracy of power load pattern recognition, a power load pattern recognition model based on improved particle swarm optimization（IPSO）algorithm to optimize long short-term memory（LSTM）neural network is proposed. By introducing diversified initial parameters, dynamic nonlinear weights and elimination mechanism, the optimization ability of PSO algorithm is improved, the key parameters of LSTM are optimized, and optimal parameter combination of LSTM neural network is determined. Experimental results show that this method can effectively improve the accuracy of the model and save the training time of the model.