Accurate multi-step traffic flow prediction plays a crucial role in intelligent transportation systems, enabling efficient traffic management and congestion mitigation. However, existing sequence-to-sequence models often fail to capture dynamic spatiotemporal dependencies, particularly struggling with evolving traffic network topologies and long-range temporal correlations. To tackle these challenges, we propose a sequence-to-sequence model based on Reinforced Attention Graph Convolutional Networks to comprehensively utilize the spatiotemporal dependencies in traffic data. Specifically, we integrate Graph Convolutional Networks with Gated Recurrent Units to enhance spatiotemporal feature extraction capabilities. On the decoder side, we propose an exponential-decay temporal attention mechanism that effectively captures features from distant time steps. Additionally, we develop an input-feedback-based training strategy to boost the model’s capacity to retain historical attention information, effectively mitigating error accumulation and enhancing prediction accuracy. Extensive experiments on two real-world traffic datasets, SZ-taxi and Los-Loop, confirm the superiority of the proposed model over existing baseline methods in multi-step traffic flow prediction.