Traffic flow prediction is a fundamental problem for urban traffic control and management. In practice, incomplete data is a common challenge due to sparse sensor deployment, data loss, and hardware failure. In this paper, a graph convolution recurrent neural network is proposed for traffic flow prediction, with considerations of incomplete data. The missing data is complemented with node imputation using the Gaussians Mixture Model (GMM) and integrated into the initial layer of the graph convolution network. Then, we utilize the node parameter learning module to capture the features of individual nodes, and the node-embedding matrix is applied to balance the computational efficiency and model performance. In addition, we employ recurrent neural networks and Sequence to Sequence models to tackle the challenge of temporal dependence and multi-step prediction. The proposed approach is evaluated based on two real-world datasets, and the results show that the prediction accuracy can be improved by at least 12.5% and 18.6% compared to the imputation and inductive-based models.