Short-term traffic flow prediction is the key for Intelligent Transport Systems (ITS) to achieve traffic control and traffic guidance． Traditional convolutional neural networks are difficult to obtain long-term memory in short-term traffic flow prediction，and there is information “leakage”from future to past． Dilated-Causal Fully Convolutional Networks (DCFCN)，which introduces dilated convolutions to increase the reception field and ob-tain long term memory of sequences，was proposed in this study． At the same time，the causal convolutional was introduced into DCFCN to solve the problem of information“leakage”． DCFCN is composed of 6 convolutional la-yers，each layer achieves causal convolution by padding，and the dilated increases exponentially with the depth of the network． Experiments show that the proposed DCFCN is superior to other comparison models in short-term tra-ffic flow prediction，and the computational efficiency on the GPU is significantly improved．