Abstract: To predict the electricity consumption demand accurately, based on the data set of PJM company in the United States, the deep autoregressive recurrent networks（Deep AR） time-series model was utilized to predict the electricity consumption of Commonwealth Edison Company at a certain 12-hour interval in the future. Based on the distribution parameters of the data in the long short term memory network（LSTM）, the predicted value was obtained by sampling in the distribution. Mean absolute error（MAE）, root mean square error（RMSE） and mean absolute percentage error（MAPE） were used as evaluation indexes for predicting short-term electricity consumption, and the model was compared with the time-series model of autoregressive integrated moving average（ARIMA） algorithm model and the Prophet algorithm model. The results show that the three performance indexes of MAE、RMSE and MAPE of the Deep AR algorithm model in predicting short-term electricity consumption are respective 1 070.01, 1 279.31 and 6.12% with high prediction accuracy. The proposed algorithm can not only predict electricity consumption in the future, but also can predict the probability distribution for further describing the globality of events.