Accurate traffic prediction under various conditions is an important but challenging task. Due to the complicated non-stationary temporal dynamics in traffic flow time series and spatial dependencies on roadway networks, there is no particular method that is clearly superior to all others. Here, we focus on investigating ensemble learning that benefits from multiple base models, and propose a traffic-condition-aware ensemble approach that acts as a decision maker by stacking multiple predictions based on dynamic traffic conditions. To sense traffic conditions, we apply the Convolutional Neural Network (CNN) model to capture the spatiotemporal patterns embedded in traffic flow. Then, the high-level features extracted by CNN are used to generate weights to ensemble multiple predictions of different models. Extensive experiments are performed with a real traffic dataset from the Caltrans Performance Measurement System. We compare the proposed approach with competitive models, including Gradient Boosting Regression Tree (GBRT) model, Weight Regression model, Support Vector Regression (SVR) model, Long Short-term Memory (LSTM) model, Historical Average (HA) model and CNN model. Experimental results demonstrate that our method can effectively improve the performances of traffic flow prediction.