•Energy consumption patterns classification was applied to building energy prediction.•Four energy consumption patterns were classified according to the analysis of decision tree.•Ensemble learning models were established to predict hourly energy consumption based on four patterns.•The value of proposed method was evaluated in different training cases. Accurate building energy consumption prediction plays an important role in building energy management and energy policy. However, traditional prediction methods of building energy consumption fail to consider the running conditions of buildings in different periods, which results in the failure of best forecasting effect. This study presents a prediction strategy of building energy consumption based on ensemble learning and energy consumption patternclassification. Hourly meteorological data from a meteorological station and energy consumption data from an office building in New York City are used for this work. First, decision tree is employed to mining energy consumption patterns and classify energy consumption data into corresponding categories. Then, the ensemble learning method is employed to establish energy consumption prediction models for each pattern. Finally, the prediction accuracy of the proposed method is compared with other three methods, i.e., ensemble learning without energy consumption pattern classification, SVR and ANN. Also, the robustness of various methods is investigated by comparing their prediction performance under different training data amounts. Results show that there are four classified energy consumption patterns of the building and significant differences among them. The ensemble learning model with energy consumption pattern classification achieves the best prediction with 17.7%, 16.1%, 15.4%, 15.8%, 15.6% of CVRMSE under 20%, 40%, 60%, 80% and 100% data availability, respectively. It illustrates that the proposed strategy is reliable and effective. Additionally, this strategy can obtain acceptable performance with less training data, which is helpful to the application of energy consumption prediction.