Short-term origin-destination demand forecasting in rail transit systems: parallel model architecture and gravity approach

Short-term origin-destination (OD) forecasting in passenger rail transit is notoriously difficult due to the high magnitude of scale, dimension, noise, and skewness of the OD matrices. Additionally, unforeseen OD data remains an issue for the data-driven models. In this paper, we address the issue of unforeseen OD data in real-time forecasting case, by using the adjusted parallel model architecture (APMA), a forecasting improvement strategy, and the reconstruction of the problem into an origin-based vector sum projection gravity (OVG) model. The forecasting problem can be split into concatenation and chained forecasting scenarios. Variations of the APMA model are tested on real-time datasets from Bangkok Subway. The proposed APMA-OVG model performs satisfactorily when compared to recent benchmarks for the unforeseen data on the concatenation case. The performance deteriorates in the chained forecast case due to accumulated error of the step-wise forecasting process when forecasting steps are broken down.