Abstract: Three general input-output prediction models: back propagation neural network (BPNN), classification and regression tree (CART) and support vector regression (SVR) are established to predict the energy consumption of subway station. The data mining algorithm is used to improve the three models and the prediction results of them based on time delay are obtained. Through comparing the results before and after the improvement, the optimal time delay is determined. Results show that among the general input-output models, the prediction of SVR model is the most accurate in terms of the energy consumption. The energy consumption prediction model based on time series contributes to the maximum improvement of BPNN prediction model. When the time delay is 5 min, the three models could achieve the best prediction accuracy, but the time series prediction model based on CART is the most sensitive one to time delay.