Abstract: In the short-term power demand forecasting research work, due to the short-term power demand being affected by many complex non-linear factors, the problems of insufficient feature consideration and the inability to model the time series correlation have appeared. This paper proposes a short-term power demand forecasting method based on characteristic analysis of long and short-term memory neural network.According to the characteristics of power demand, the influencing factors are analyzed and mined, and then the network model combining linear regression and long short-term memory neural network is used to model the power demand. Through the feature engineering, the correlation between each influencing factor and the final result benchmark and sensitive quantity is mined, and the problem that the feature expression ability is not strong enough is solved. In the neural network modeling, the structure design of forgetting gate, input gate and output gate is adopted to enhance the learning ability of the model to the sequence correlation. In this paper, we choose the 2015-2020 power data of a province to verify the proposed method, and obtain higher precision forecasting results, which proves that the proposed method can improve the modeling results and improve the accuracy of short-term power demand forecasting.