To solve the problem of low accuracy of load prediction caused by the interaction of multiple factors, a new method of long short-term memory recursive neural network (LSTM) short-term load prediction based on feature correlation analysis correction and global particle swarm optimization (GPSO) was proposed. This method firstly carries on the exploratory data analysis (EDA) and preprocessing to the load correlation sequences, finds the intrinsic mechanism of the characteristics and the correlation relations, and then revises them to ensure the strong correlation and integrity of the input characteristics. Aiming at the defects of the traditional feedforward neural network which cannot process the sequence correlation information and the ordinary recuisive neural network which cannot remember the remote key information, Deep learning based on LSTM load prediction model was constructed. Due to the random initialization of LSTM network weights, the objective function is prone to fall into the local optimal during the training process. The improved particle swarm optimization algorithm is used to optimize the network weights of the prediction model and improve the overall prediction ability of the model. Compared with the benchmark models of back propagation neural network (BPNN) and recursive neural network (Elman), the prediction accuracy of the proposed model method is significantly improved.