MetroFlowNet: Adaptive Multi-Feature Fusion for Metro Passenger Movement Prediction

Metro systems generate passenger demand that changes across time, stations, travel directions, and operat ing conditions. Reliable passenger-flow prediction can support crowd management, service planning, and data-driven opera tional decisions. This paper presents MetroFlowNet, a project specific framework for predicting passenger movement from historical smart-card tap-in/tap-out observations. The frame work represents passenger demand through station-level and origin–destination (OD) information and combines temporal, spatial, and contextual variables before sequence prediction. The design emphasizes adaptive feature fusion so that the contribution of different feature groups can vary with the observed operating state. The proposed workflow contains data cleaning, time window aggregation, OD construction, temporal feature genera tion, spatial representation, feature normalization, model train ing, and multi-horizon evaluation. Recent research shows that passenger-flow forecasting benefits from models that represent both temporal dynamics and relationships among stations, while decomposition and multi-time representations can improve the treatment of recurrent demand patterns [1]–[5]. MetroFlowNet is designed as an independently implemented framework rather than a reproduction of any single published architecture. The final experimental study will compare MetroFlowNet with se lected baselines using MAE and RMSE and will include ablation experiments to quantify the contribution of each feature group. Project-specific dataset statistics and measured test results are intentionally reserved for insertion after the final experimental run.