Designing a Novel Two-Stage Fusion Framework to Predict Short-Term Origin–Destination Flow

Abstract

Short-term origin-destination (OD) demand predicting plays an indispensable role in intelligent transportation systems and ride-hailing service operations. However, most studies are carried out in trip ends travel demand prediction, i.e., trip generation/attraction, while paying less attention to predicting OD flows in citywide traffic management. To this end, a novel generalized framework is proposed, named two-stage fusion framework (TFF). TFF consists of two stages, where the first stage utilizes an attention-based spatio-temporal graph convolutional network (AST-GCN) to predict the trip generation/attraction, and the second stage develops a modified Kalman filter (KF) to predict OD flow which is converted into a coefficient matrix. Lastly, the final predicted OD can be obtained by integrating the trip generation/attraction and the coefficient matrix. In AST-GCN, a gated fusion mechanism and dynamical zone proximity matrix are applied to improve the capacity of capturing the spatio-temporal interdependence among traffic analysis zones. In KF, we redefine the state vector and observed vector and their covariance matrixes, and then introduce the Box-Cox technique to standardize the deviation matrix to adapt to KF. The proposed model is tested on a large-scale real-world data set from Chongqing, China, and the results indicate that TFF can achieve a satisfactory performance of short-term OD prediction. Moreover, the proposed framework has the convenient scalability, because either of the two stages is replaceable by other predictors.