Accurate traffic flow prediction plays a critical role in intelligent transportation systems, providing essential support for urban planning, traffic control, and congestion mitigation. To address the challenges of spatial heterogeneity and temporal dynamics inherent in traffic data, this paper proposes AST-GNNFormer, an adaptive spatio-temporal graph neural network that integrates graph attention mechanisms with temporal convolution. The model introduces three key components to enhance predictive accuracy and generalization: (1) a Layer-aware Information Preservation mechanism that mitigates over-smoothing in deep GNNs by retaining original node features across layers; (2) an Inter-Layer Attention Module that dynamically selects and weights informative layer-wise features to improve multi-layer fusion quality; and (3) an Adaptive Graph Learning Module that fuses prior adjacency knowledge with learnable structures, enabling dynamic topology adaptation. Additionally, a Temporal Convolution Module is incorporated to model multi-scale temporal dependencies efficiently. Extensive experiments on real-world benchmark datasets (PEMS04 and PEMS08) demonstrate that AST-GNNFormer significantly outperforms existing state-of-the-art methods in both short-term and long-term traffic forecasting tasks. Ablation studies further confirm the effectiveness of each proposed component.