The spatial-temporal forecasting has attracted remarkable attention due to the increasing number of widespread applications with electricity, finance, climate, environment, and traffic. Although various studies on spatial-temporal forecasting exist, these algorithms leverage the same neural network to model all traffic series, which unreasonably and incapably learn spatial-temporal correlations from traffic series with different patterns using common parameters, and unfortunately, none of these algorithms considers data loss of time series, which greatly affects forecasting performance. Due to such, we propose a novel Cluster and Completion Graph Convolution Recurrent Network (CCGCRN) for incomplete traffic flow forecasting. To learn the spatial-temporal correlations better, we utilize CCGCRN to learn the common parameters within the cluster with strong relevance among traffic flows. Three techniques are proposed: (1) To learn the common parameters for traffic series sharing the same patterns, we propose the Cluster Parameter Learning module (CPL) to cluster traffic series based on the orthogonal non-negative matrix factorization algorithm, (2) To complete the missing values of traffic series by utilizing the commonality of series within a cluster, we propose the Completion Learning Network (CLN) to learn the spatial-temporal interaction relationships, and (3) To accurately forecast traffic series, we propose the Graph Convolution Recurrent Neural Network (GCRNN) that learns the graph construction adaptively and captures spatial-temporal correlations in each cluster. Extensive experimentations on two real-world datasets with 10% and 30% missing rates demonstrate that the proposed model achieves superior forecasting performance than the other eight baselines.