Abstract
<title>Abstract</title> <p>Accurate short-term prediction of inbound passenger flow for urban rail transit is of great significance for network planning, smart services, and refined operations. However, existing methods cannot deeply fuse spatial topology, temporal periodicity, and trends, and fail to capture dynamic fluctuations and long-term evolution, causing low accuracy. To address this, we propose AGHA-Net, a method based on adaptive graph convolution and hybrid attention. The method first uses an adaptive graph convolutional network (AdaGCN) to accurately capture dynamically evolving spatial dependencies among stations. Then, a spatiotemporal gated recurrent unit (ST-GRU) deeply couples graph convolution with temporal recurrence, enabling efficient fusion and extraction of local spatiotemporal features. Furthermore, a dual-branch hybrid attention layer is introduced that uses sequential self-attention to capture long-range temporal dependencies. In contrast, a bilinear temporal attention mechanism performs dual-dimensional compression and constructs a temporal attention matrix to model internal correlations within long sequences. Consequently, it comprehensively improves prediction accuracy and stability. Experiments on Hangzhou metro data show AGHA-Net significantly outperforms baselines. Visualization verifies its ability to capture complex patterns, and ablation studies confirm each module’s necessity. The proposed method provides reliable decision support for intelligent rail transit management, passenger flow control, and capacity optimization.</p>