This paper presents a data mining (DM) based approach to developing ensemble models for predicting next-day energy consumption and peak power demand, with the aim of improving the prediction accuracy. This approach mainly consists of three steps. Firstly, outlier detection, which merges feature extraction, clustering analysis, and the generalized extreme studentized deviate (GESD), is performed to remove the abnormal daily energy consumption profiles. Secondly, the recursive feature elimination (RFE), an embedded variable selection method, is applied to select the optimal inputs to the base prediction models developed separately using eight popular predictive algorithms. The parameters of each model are then obtained through leave-group-out cross validation (LGOCV). Finally, the ensemble model is developed and the weights of the eight predictive models are optimized using genetic algorithm (GA). The approach is adopted to analyze the large energy consumption data of the tallest building in Hong Kong. The prediction accuracies of the ensemble models measured by mean absolute percentage error (MAPE) are 2.32% and 2.85% for the next-day energy consumption and peak power demand respectively, which are evidently higher than those of individual base models. The results also show that the outlier detection method is effective in identifying the abnormal daily energy consumption profiles. The RFE process can significantly reduce the computation load while enhancing the model performance. The ensemble models are valuable for developing strategies of fault detection and diagnosis, operation optimization and interactions between buildings and smart grid. (c) 2014 Elsevier Ltd. All rights reserved.