Short-term passenger flow forecasting method of rail transit under station closure considering spatio-temporal modification

To realize the accurate prediction of the short-term passenger flow of rail transit and explore the changing mechanism of passenger flow under the station closure, a short-term spatio-temporal corrected passenger flow forecasting method considering dynamic factor model (DFM) and support vector machine (SVM) under the station closure was developed and denoted by DFM-SVM. A hybrid model combining symbolic aggregation approximation (SAX) and dynamic time warping (DTW) denoted by SAX-DTW was proposed to identify the spatio-temporal ranges of the affected stations. DFM was developed to forecast the short-term passenger flow under the normal scenario based on the historical data. SVM was developed to extract and process the nonlinear characteristics of the passenger flows at the affected stations and time periods and used to correct the correspondingly affected passenger flows. The validity of the method was verified by an example of the inbound volume prediction at the Beijing Subway Station under the station closure. Research results show that compared with the SAX, the proposed SAX-DFM not only comprehensively considers the changes in the number and trend of passenger flow, but also identifies the abnormal segments of several stations according to the case study more accurately. Compared with the traditional DFM, the proposed DFM-SVM can significantly reduce the forecasting residual errors of passenger flows at each station. Taking the Olympic Sports Center Station as an example, the residual error reduces by about 60%. In terms of overall passenger flow prediction of the whole stations, the proposed DFM-SVM reduces the root mean square errors by 43.39%, 70.00%, 33.18% and 70.83%, respectively, and the mean absolute errors by 43.72%, 67.17%, 28.98% and 57.08%, respectively, compared with the baseline models such as Holt-Winters, SVM, gate recurrent unit (GRU), and long short-term memory (LSTM). In terms of the passenger volume prediction at a single station, the proposed DFM-SVM can reduce the root mean square errors and mean absolute errors at about 70% stations compared with other benchmark models. Therefore, the proposed DFM-SVM can capture the nonlinear feature of passenger flow affected by the station closure, which greatly improves the prediction accuracy and provides reliable passenger flow's early warning information and decision-making basis for operation managers. 4 tabs, 9 figs, 30 refs.