Abstract: Addressing the deficiencies of existing deep learning models in local spatio-temporal dependency modeling and local-global feature fusion, a new prediction model LFAGCSA (Local Fusion Adaptive Graph Convolution And Self-Attention) is proposed. The model achieves refined modeling through a multilevel architecture: (1) temporal, feature, and spatial embedding techniques are used in the data processing stage to map the original data into a high-dimensional representation space to enhance feature separability; (2) dynamic adaptive graph convolution and gated deep convolutional feedforward network (GDFN) are combined in the local feature extraction module to achieve adaptive modeling of the local road network structure; (3) temporal and spatial self-attention mechanisms are integrated in the global feature capture module to model the global dependencies across time and regions, respectively; and (4) the local fine-grained features are organically combined with the global spatial-temporal features in the fusion output layer, and ultimately outputting the multistep prediction results through the fully connected layer. The experimental results on four real datasets demonstrate that LFAGCSA outperforms the benchmark model in all evaluation criteria, achieving accurate prediction of future traffic flow.