A Study on Dynamic Estimation and Short-Term Prediction of Urban Rail Transit Passenger Flow OD Using Multi-source Data

A lag in the formation of actual Origin-Destination (OD) passenger flow data in urban rail transit systems prevents short-term OD prediction from directly adopting previous actual OD data as input, compromising prediction timeliness and accuracy. To address this issue, this paper proposes an integrated “OD dynamic estimation - short-term OD prediction” scheme and constructs a Computational Graph-Long Short-Term Memory (CG-LSTM) hybrid prediction model. Based on the spatiotemporal evolution law of network passenger flow, the model integrates multi-source real-time data (e.g., from Automatic Fare Collection (AFC) systems and train weighing equipment) to build a hierarchical computational graph network of “entry - OD - path - section - exit”. Through the iteration of “forward generation - deviation calculation - reverse correction”, it achieves dynamic estimation of current and previous-period OD. Subsequently, with the estimation results and historical of the same period OD data as input, the LSTM component captures temporal features and outputs OD predictions for the next 15 min. A case study on a local Guangzhou Metro network verifies the scheme’s effectiveness: for flows with OD volume >10 (15-min granularity), the prediction error is 0.117. This research provides support for the intelligent operational decision-making of urban rail transit.