Traffic flow forecasting is an important problem of data mining and has widespread real-world applications. To model the complex spatial-temporal dependencies in traffic data, various models such as spatial-temporal graph convolutional networks (STGCNs) have been proposed and achieved advanced performance. However, we have identified the following limitations in these methods: i) Most methods fail to consider the spatial-temporal indistinguishability of samples or only distinguish samples at the input end, leaving the samples indistinguishable in the deeper layers of the network. ii) Most GCN-based methods use the same adjacency matrix at every layer of the model, which restricts the ability to capture deep semantic information. In this paper, we propose a new model called IDMixer to address these limitations. Specifically, unlike most STGCNs that use GCNs to propagate temporal information between nodes, we use GCNs to extract the spatial-temporal identities of nodes. Furthermore, since traffic data may exhibit different dependencies in seasonal and trend components, we decompose the traffic data into seasonal and trend components at each layer and then extract their spatial-temporal identities separately. The identities and learned spatial-temporal features are then mixed by Multi-Layer Perceptrons (MLPs). Additionally, we construct a semantic adjacency matrix based on the similarities between node features to help the GCNs capture deep semantic information. Experiments on six real-world traffic datasets demonstrate that our model outperforms state-of-the-art models.