LSTM-TCN Forecasting for Short-Term Passenger Flow at Integrated Transport Hubs Using a Population-Mobility Proxy: Ablation Evidence from Shenzhen North Station

Abstract

Short-term passenger-flow forecasting at integrated transport hubs requires accurate forecasts with explicit target semantics. This study evaluates long short-term memory (LSTM) and temporal convolutional network (TCN) fusion models using two distinct 5 min targets at Shenzhen North Station: a 2023–2026 regional population-mobility crowd-density proxy and a 2025 automatic fare collection (AFC) exit-card throughput series. On the proxy task, chronological 70/15/15 splits, a 12-step purge and full-epoch neural training showed that the LSTM-only variant (M2) achieved the lowest neural mean absolute error (MAE) at 5 min, while the fused LSTM–TCN variant (M4) achieved the lowest neural MAE at 15–60 min. Ridge was the strongest conventional proxy reference at 5 min and Random Forest at 15–60 min. On AFC exits, Random Forest was strongest among the six conventional references at all four horizons, while M4 remained better than ridge at 5–60 min. The raw 5 min proxy–AFC Pearson correlation was 0.553; after weekday-by-clock de-seasonalization it was 0.505 (95% day-block confidence interval 0.457–0.538). These values support a shared-signal interpretation; calibration requires a separate mapping model. In the observed-arrival pressure test, recovery at 21:10 reduced cumulative queueing by 67.3%; delays of 10 and 15 min reduced the benefit to 63.4% and 61.7%. The workflow provides 5–60 min lead time for gate and queue preparation, staff rostering and feeder coordination.