Predicting urban traffic flow is of big significant to traffic management and public security. However, with the continuous expansion of urban areas and the development of data acquisition technology, new types of traffic data are characterized by wide spatial distribution, high timeliness and large data volume. Traffic flow forecasting requires high cost and related domain knowledge. Therefore, it has become an urgent research topic to properly use a small amount of traffic data to efficiently construct a traffic prediction model. In this paper, we propose a generative adversarial network-based traffic flow prediction method called RT-GAN which is the real-time prediction of traffic flows in the surroundings area according to the traffic information in the central area. The combination of gated convolution and dilated convolution can capture the traffic information in the near and far regions and perform feature fusion to achieve real-time prediction. Experiments on the Beijing and New York traffic flow data sets show that our method outperforms others, providing a potential solution to practical applications.