Elman neural network is widely used for dynamic data prediction because of its ability to approximate and adapting to time-varying characteristics. There are many uncertain factors in the short-term electricity consumption. In order to take all the factors into account, this paper introduces the reconstruction phase space technology of chaotic time series. Due to the large deviation from the neural network of the peak prediction in nonlinear functions, it can be modified by rough set theory. Therefore, the chaotic time series theory and rough set theory are introduced to improve the Elman neural network. The model applies embedded dimension and delay time to reconstruct the phase space to restore the original system's dynamic morphology. The processed data is brought into the Elman neural network to predict the electricity consumption. Finally, the peak point corrected by the rough set is introduced to improve the prediction accuracy. This paper collects the data from a dormitory building in Heriot-Watt university of Edinburgh. It uses thirty days electricity data with 8 640 points as the data set to do predict simulation. The prediction results are compared with the Elman neural network and chaotic time series Elman neural network, and the validity of the model are verified in a short-time prediction. This work is supported by National Natural Science Foundation of China (No. 71572126).