To enhance the spatio-temporal modeling capability of traffic flow prediction, a prediction method based on an adaptive graph attention gated recurrent unit network (AGA-GRU) was proposed. The method constructed a learnable adjacency matrix and obtains a probabilistic structure through softplus activation and row normalization. A row-wise Top-K retention and renormalization strategy was introduced to achieve hard sparsification, while an L1 sparsity regularization was applied to the structural parameters to improve interpretability and robustness. In the spatial dimension, multi-head graph attention (GAT) was employed to suppress noisy edges and highlight key correlations. In the temporal dimension, gate recurrent unit (GRU) was used to capture sequential dependencies. The two components were cascaded to form an efficient spatio-temporal modeling framework. The training objective combined prediction error and sparsity regularization. Experiments conducted on real-world datasets, with autoregressive integrated moving average model (ARIMA), bidirectional long short-term memory network (Bi-LSTM), graph neural network-long short-term memory network (GNN-LSTM), spatio-temporal graph convolutional network (ST-GCN), and multi-task graph neural network (MTGNN) as baseline models, were evaluated using mean absolute error (MAE), mean squared error (MSE), and root mean squared error (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.