The importance of short-term power load forecasting continues to improve with the development of power enterprises. Although the traditional load forecast has been developed relatively mature, the load forecast accuracy requirements gradually increase now. To meet the development needs, existing methods should be improved or new methods of prediction should be established. This paper analyzes the periodic load forecasting and periodic characteristics of load forecasting data. It also combines recursive neural networks with Limited Boltzmann machine's strong unsupervised learning ability in analyzing the unique advantages of time series data. The network principle and training process of the combination are described and the electric load data are predicted by experiment. The experiment compares the accuracy of the network and other networks in short-term load forecasting. The results show that the neural network proposed in this paper is more accurate than other networks in the power short-term load prediction experiment. This work is supported by National Natural Science Foundation of China (No. 51677072).