Multi-Stage Fusion Framework for Short-Term Passenger Flow Forecasting in Urban Rail Transit Systems Using Multi-Source Data

To improve real-time operation and management in urban rail transit (URT) systems, accurate and reliable short-term passenger flow forecasting at the network level is a crucial task. Although numerous endeavors have been devoted to this field, the insufficient topological representation for passenger flows in the URT network, the overlooking of intrinsic correlations among multi-source data, and the information loss in deep-learning frameworks are still critical issues that need to be addressed. This study proposes a multi-stage fusion passenger forecasting (MSFPF) model to accomplish short-term multi-step passenger forecasting leveraging multi-source data, and overcome the above-mentioned challenges. Based on the characteristics of passenger flows in the URT network, time-based origin–destination flow data is involved and utilized to enhance the representation of flows and provide spatial-temporal features. Then, the interaction and relationship among multi-source data are estimated to capture their intrinsic correlations. To effectively and comprehensively extract temporal and spatial features, a transformer long short-term memory block and a depth-wise attention block are constructed with attention mechanisms and employed. Furthermore, we construct the multi-stage fusion (MSF) structure to alleviate the information loss during the learning process, which is a significant component in improving the forecasting accuracy. In addition, the model is applied to two large-scale real-world datasets, in which it outperforms nine widely used baselines and four specific variants of itself. The quantitative experiments demonstrate the robustness and superiority of the proposed MSFPF model, and the significant contribution of the MSF structure in the model.