With the rapid development of online ride-hailing services, people have increasingly relied on platforms providing these services to travel. The corresponding companies need to accurately obtain passengers' travel demand to allocate orders and drivers to regions. Therefore, traffic demand prediction is a critical problem of Intelligent Transportation Systems (ITS). Origin-Destination Matrix Prediction (ODMP) is a challenging extension of traffic demand prediction that needs to consider the temporal and spatial dependence of traffic data and predict the relationship between origin and destination of passengers' demand. In this paper, we proposed a method to convert order paths of passenger demand to the hexagon-based path graph. The path graph shows the origin and the destination of the paths of a period. Specifically, considering that traffic flows are time-varying, we generate different hexagon-based path graphs for different time periods. Then, we propose a Hexagon-based Dynamic-Graph Convolutional Network (Hex D-GCN) to make the GCN suitable for dynamic graphs, in which graph connections are different in time series. Furthermore, We evaluate our model on the Didi Chuxing KDD CUP 2020 dataset and get the state-of-art performance. It is shown that our method combines the spatial correlation and temporal correlation well and also captures the passenger's demand pattern.