At present, in the cruise mode, the supply and demand between taxis and passengers is not easy to match, resulting in the coexistence of empty taxis and hard hailing phenomenon. Accurate and efficient realization of the road network taxi demand forecast is conducive to effectively alleviate this problem. In the view of that the existing traffic flow prediction models can not sufficiently extract spatial features, especially the spatial relationship between sections in urban road network, full consideration was given to the three kinds of spatial relationship between segments within the road network, and three types of graph were constructed correspondingly, i.e. the local relationship graph, the global relationship graph and the OD frequency relationship graph. This paper proposes a taxi demand prediction model which is composed of graph convolutional network and temporal convolutional network. The graph convolutional network is used to mine the spatial relationship features of sections in the urban road network, and the temporal convolutional network is used to mine the time series features of traffic data, and the influence of external factors is considered. In the experiment, the number of taxi trips in each section of the urban road network is extracted from the real taxi GPS trajectory data, and the series of trips number formed in multiple time slots on the road is used to verify the prediction model. The results show that the proposed model is superior to the common traffic prediction models, and has smaller mean absolute error, root mean square error and mean absolute percentage error.