Accurate traffic flow prediction under abnormal conditions, such as accidents, adverse weather, work zones, and holidays, is significant for proactive traffic control. Here, the authors focus on a special challenge of how to develop robust responsive algorithms and prediction models for short-term traffic forecasting in different traffic conditions. To this end, this study presents an ensemble learning algorithm for the short-term traffic flow prediction via the integration of gradient boosting regression trees (GBRT) and the least absolute shrinkage and selection operator (Lasso). Four different model structures whether considering the feature selection are proposed and tested for multi-step-ahead prediction under both normal and abnormal conditions. The results indicate that the proposed multi-model ensemble models are superior to the benchmark algorithms, i.e., support vector regression, and random forests, the GBRT model outperforms the Lasso model under normal traffic conditions, and the Lasso model has a better prediction accuracy under abnormal traffic conditions. In addition, the Lasso model with the feature selection is superior to the full feature model under either normal or abnormal conditions, while the GBRT model is not always better under normal conditions. The proposed integration framework is general and flexible to assemble various traffic prediction algorithms.