With the rapid development of Intelligent Transportation Systems (ITS), accurate short-term traffic flow prediction is becoming more and more vital for both traffic management authorities and residents, since it can improve traffic management efficiency and personal travel experience. However, many existing prediction methods rely solely on traffic flow data, neglecting the effects of some external factors (e.g., specific dates and actual road conditions on traffic flow), which limit to achieve accurate prediction. Therefore, this paper proposes a deep learning-based AT-Conv-GRU (Attention Convolutional-Gated Recurrent Unit) model, which takes historical traffic flow data and external factors as input data simultaneously, and takes the traffic flow at predicted moment as output. Firstly, the historical traffic flow data of the current moment is input into the Conv-GRU (Convolutional-Gated Recurrent Unit) module with attention mechanism to extract features from both spatial and temporal dimensions. The historical traffic flow data of the same moment of the previous day and the last week is input into the Bi-GRU (Bi-directional GRU) module to extract time-dependent and periodic features. Meanwhile, the hidden features of external factors are extracted using CNN (Convolutional Neural Network). Finally, the extracted features of each module are concatenated by feature fusion layer and eventually output by fully connected layer. Extensive experimental results show that the prediction accuracy of the AT-Conv-GRU model significantly outperforms traditional methods and other deep learning-based methods.