Load forecasting is the basis of safe operation of power system. Due to the randomness and volatility of residential electricity load, it may affect the normal operation and maintenance of power system. Therefore, accurate prediction of residential power load provides favorable guidance for real-time dispatching of power grid. In this paper, an ultra-short-term prediction method for residential electricity load based on long-short-time memory-type cyclic neural network is proposed. The “memory” feature of this method is used to mine the correlation characteristics between load data, and a resident based on long-short-term memory network is established. The ultra-short-term prediction model of electric load is compared with the simulation results of the double-layer feedforward neural network model. The prediction results based on the long-short-time memory network are more accurate, and the validity of the model is verified.