To address the insufficient fusion of multi-granularity temporal features and inadequate modeling of spatial heterogeneity among stations in short-term metro inbound passenger flow forecasting, this paper proposes the Multi-Granularity and Attention-enhanced Passenger Flow Prediction model (MGAPF). In the temporal dimension, four one-dimensional convolutional branches with different receptive fields are constructed, and a granularity attention mechanism is used to adaptively fuse multi-scale temporal features. In the spatial dimension, a graph attention mechanism is introduced to dynamically learn the correlation strengths among adjacent stations. External temporal contextual features, including holidays, weekends, and weather conditions, are also incorporated. A gated fusion structure, residual connections, and a Transformer encoder are further combined to perform spatiotemporal feature interaction and generate prediction outputs. Experimental results based on real operational data from Hangzhou Metro show that MGAPF achieves MAPE values of 16.93% and 9.52% in the 10-minute and 30-minute forecasting tasks, respectively, representing error reductions of approximately 4.5% and 7.8% compared with the second-best baseline model. MGAPF also performs better in terms of MAE, RMSE, and prediction stability, indicating that the fusion of multi-granularity temporal features, spatial correlations among stations, and external factors can effectively improve the accuracy and stability of short-term metro inbound passenger flow prediction.