Short-term cross-sectional passenger-flow forecasting for urban rail transit with multi-attribute feature fusion

Accurate and reliable short-term cross-sectional passenger-flow forecasting is essential for optimizing operation strategies and advancing intelligent urban-transport management. However, due to the combined effects of weather, calendar attributes, and other external factors, existing short-term cross-sectional forecasting models still struggle to fully capture complex influencing mechanisms with high accuracy. To improve forecasting accuracy, we first construct an inter-station trip-gravity indicator using the gravity model and incorporates it with weather and calendar attributes to form a multi-attribute passenger-flow feature vector. A gated recurrent unit (GRU) backbone is employed to extract temporal dependencies of cross-sectional flows. Meanwhile, a Transformer equipped with multi-head self-attention is used to capture non-linear interactions between exogenous factors and the GRU-derived temporal features with attention weights learned in an end-to-end manner. Hyper-parameters, including network depth, batch size, learning rate and training epochs, are globally optimised via particle-swarm optimisation, while a fully-connected layer performs feature fusion and final prediction. Using passenger-flow data from Hangzhou Metro Line 1 in January 2019 as an example, experimental results show that the proposed method achieves mean absolute percentage errors (MAPE) of 14.96%, 15.86% and 16.96% at 10, 15 and 30 minutes intervals, respectively, outperforming benchmark models by a notable margin.