Adaptive Graph Learning with Persona-Aware Attribution for Event-Resilient Metro Passenger Flow Prediction

Abstract

Urban rail passenger flow forecasting, based on large-scale automatic fare collection (AFC) sensor networks, is typically evaluated on ordinary days, yet operators most need accurate predictions during event days. The AFC infrastructure comprises 191 sensor-equipped stations with turnstile transducers that generate 95.7 million discrete sensing events (inboard/outboard readings) over 57 days, constituting a high-velocity, multimodal spatiotemporal sensor stream. This study presents a topology-driven adaptive graph network (TDAG-Net) for event-day forecasting across a full metro system. The model learns spatial dependencies end-to-end by merging multi-scale temporal convolution with an adaptive graph attention branch through learned gates. An attention long short-term memory (LSTM) then decodes 60-min forecasts for all 191 stations at once. A ticket–persona module reads 25 fare channels as four rider types to decompose demand surges by traveler profile. The proposed model reduces root mean square error (RMSE) by 9.29% and achieves the best performance on 22 flagged event days. Ablation reveals that adaptive adjacency is the dominant component and reduces horizon decay to 17.3% from 33.0%. Counter to expectation, persona-specific graphs are 98.55% identical, and splitting inputs by rider type degrades accuracy by 10.36% RMSE. On event peaks, the tourist/visitor share rises from 5.7% to 16.5%, indicating that event-day crowd management should prioritize unfamiliar riders over regular commuters.