To enhance the spatio-temporal modeling capability of traffic flow prediction, this paper proposes a prediction method based on an Adaptive Graph Attention Gated Recurrent Unit network (AGA-GRU). The method constructs a learnable adjacency matrix and obtains a probabilistic structure through softplus activation and row normalization. A row-wise Top-K retention and renormalization strategy is introduced to achieve hard sparsification, while an L1 sparsity regularization is applied to the structural parameters to improve interpretability and robustness. In the spatial dimension, multi-head Graph Attention (GAT) is employed to suppress noisy edges and highlight key correlations; in the temporal dimension, GRU is used to capture sequential dependencies. The two components are cascaded to form an efficient spatio-temporal modeling framework. The training objective combines prediction error and sparsity regularization. Experiments conducted on real-world datasets, with ARIMA, Bi-LSTM, GNN-LSTM, ST-GCN, and MTGNN as baseline models, are evaluated using MAE, MSE, and RMSE metrics. The results show that AGA-GRU outperforms the compared methods in multi-step forecasting tasks and learns sparse and interpretable graph structures (e.g., self-loops and a few key cross-link connections), achieving a good balance between prediction accuracy and model interpretability. This method is well-suited for traffic flow prediction in urban road networks.