<p>This study addresses the challenges of modeling spatiotemporal dynamics in short-term rail transit passenger flow prediction (PFP), as well as data sparsity issues caused by new stations and unexpected events. It proposes a novel prediction model, STGCN-Meta, which integrates a Spatio-Temporal Graph Convolutional Network (STGCN) with meta-learning. The model is designed to capture the nonlinear propagation patterns of passenger flow within complex transit networks. The methodology involves three key components. First, a dynamic STGCN replaces the traditional fixed adjacency matrix by introducing an attention mechanism to adaptively learn the evolving correlation weights between stations, accurately representing spatiotemporal dependencies. Second, a model-agnostic meta-learning framework is incorporated, allowing the model to rapidly adapt to new prediction tasks with minimal samples through multi-task parameter initialization. Third, a multi-source data fusion module employs an attention mechanism to integrate external factors, including weather and holidays, into the prediction process. Experimental results demonstrate that for one-hour-ahead PFP, STGCN-Meta achieves a mean absolute error of 9.54, a root mean square error of 14.26, and a mean absolute percentage error of 6.42%, consistently outperforming all baseline methods. Ablation studies and case analyses further confirm the effectiveness of each module and the model’s ability to respond to sudden changes in passenger flow. In conclusion, STGCN-Meta provides a high-precision, generalizable solution for short-term PFP. Its combination of dynamic graph learning and meta-learning mechanisms also offers a novel approach for other spatiotemporal prediction problems in transportation systems.</p>