Traffic flow prediction is an important task for intelligent transportation systems and congestion mitigation. Traditional methods mainly rely on statistical approaches, which fail to capture the complex spatiotemporal correlations within traffic flow. In recent years, graph neural networks (GNN) have shown remarkable performance in modeling and predicting traffic flow. However, existing GNN-based methods face limitations in capturing the dynamic spatiotemporal dependencies and latent interactive relationships, which hinders prediction accuracy. To address this issue, a novel method named temporal attention and interactive dynamic graph convolutional network (TAIDGCN) is proposed. Dynamic temporal dependencies are captured using a temporal attention layer (TAL) and a dilated temporal convolution layer (DTCN), while spatial dependencies are modeled through an interactive dynamic graph convolution network (IDGCN). A multi-component structure is further integrated to enhance the overall spatiotemporal modeling capability. Specifically, TAL and DTCN extract temporal features from recent patterns, daily cycles, and weekly cycles, thus improving mid- and long-term spatiotemporal prediction accuracy. Meanwhile, IDGCN leverages an interactive learning mechanism to capture dynamic spatial relationships in traffic flow. The proposed model is validated on two real-world open highway datasets. Experimental results demonstrate that TAIDGCN effectively captures dynamic spatiotemporal information and outperforms ten benchmark models in prediction accuracy, highlighting its superiority and potential for traffic flow prediction applications.