Short-term prediction of passenger volume for urban rail systems: A deep learning approach based on smart-card data

Abstract

Short-term prediction of passenger volume is a complex but critical task to urban rail companies, which desire prediction methods with high accuracy, time efficiency and good practicality. Good prediction results of the outbound passenger volume at urban rail stations are important to the organization of passenger flow, and helpful to the arrangement of shuttles, especially in large transit junctions. The application of automatic fare collection (AFC) devices in urban rail transit systems helps to collect large amounts of historical data of completed journeys, which can be used by metro operators to construct a database of the urban rail passenger volume. Based on deep learning techniques and big data, this paper develops an improved spatiotemporal long short-term memory model (Sp-LSTM) to forecast short-term outbound passenger volume at urban rail stations. The proposed model predicts the outbound passenger volume on the basis of the historical data of the spatial-temporal passenger volume, station origin–destination (OD) matrix and the operation data of the rail transit network. Finally, based on actual data of the Beijing Metro Airport Line, a case study is carried out to compare the proposed Sp-LSTM with other prediction methods, i.e., the general long short-term memory model (LSTM), the autoregressive integrated moving average model (ARIMA), and the non-linear autoregressive neural network model (NAR), and the results show that the proposed method outperforms the others.

Introduction

The construction of urban rail transits is in fast development in the beginning of the twenty first century of China. Urban railway is a safe and sustainable transportation mode with large capacity, thus plays an increasingly important role in modern cities (Kai et al., 2016; Yang et al., 2019). In China, by the end of 2018, up to 185 urban railway lines were put into operation in 35 cities, with a total mileage of 5761.4 km. The total mileage of urban railways in operation in megalopolises such as Beijing and Shanghai exceed 700 km. The instantaneous passenger volume at urban rail transit stations can be overwhelming during peak hours, under such circumstances, the problem of emergency safety of every passenger must be guaranteed in consideration of the closeness of stations (Lam et al., 1999). Therefore, good prediction results of short-term outbound passenger volume at stations not only help to organize passengers thus keep them safe, but also contribute to better arrangement of shuttles, especially in large transit hubs. Moreover, accurate short-term prediction is conducive to the appropriate staffing in stations, based on which the metro operators can organize and guide passengers appropriately in advance to avert accidents that can harm mass passengers (Hu, 2011; Li and Lo, 2014a; Yin et al., 2019). Accurate short-term prediction can also help metro operators to design efficient train timetable, and make effective operation plans, thus the travel demands of large amounts of citizens can be satisfied and the overall efficiency of the urban rail transit system can be improved (Xue et al., 2019). This paper focus on the study of the prediction of short-term outbound passenger volume at urban rail transit stations, which proposes an improved spatiotemporal long short-term memory model (Sp-LSTM) aiming to improve prediction performance and provide rational suggestions to metro operators for better management of the railway system.

Passenger demands change frequently in daily operations, which necessitate a prediction model that is robust and flexible enough (Li and Lo, 2014b). The prediction of short-term passenger volume is to estimate the passenger volume in the near future, based on the information given by the historical data. In the literature, the time span of the short-term prediction of passenger volume varies from 5 min to one week. The problem to determine proper prediction time scale is critical to the focused problem in this paper, which directly affects the accuracy of the prediction results. The time spans of prediction in this paper are selected as 15 min, 30 min and 60 min, and the prediction results under different choices of prediction time scales are discussed. Due to the application of automatic fare collection (AFC) devices in rail transit systems, metro operators are able to obtain a substantial amount of smart card data, enabling the construction of a large database that includes station inbound/outbound passenger volume, and a station origin–destination matrix (Reddy et al., 2009; Wang et al., 2019). At present, deep learning technologies, which have good abilities for learning and generalization, provide good theoretical methods to tackle large-scale data (Najafabadi et al., 2015; Qiu et al., 2016).

The inbound and outbound passenger volume at stations are measured at a fixed time interval. The long short-term memory model (LSTM) based on deep learning takes sequential data of passenger volume as input, and the output is influenced by the input and output of the model in previous time periods. Such a structure makes it possible to have a certain memory of the processed information (Gers et al., 1999). The LSTM is especially well adapted for learning time series data with long-term dependence and has been widely used in the prediction of passenger volume. However, a notable disadvantage of LSTM is that it pays close attention to data timeliness, but pays little attention to the spatiality of the data, which would greatly affect the prediction performance. Therefore, this paper proposes an improved LSTM (i.e., Sp-LSTM) based on spatiotemporal passenger volume to predict short-term passenger volume.

The remainder of this paper is organized as follows. In Section 2, we review the relevant literature and provide the problem statement. In Section 3, we elaborate on the LSTM and Sp-LSTM. Based on real-world passenger and operation data from Beijing Metro Airport Line, we conduct a case study in section 4. Section 5 concludes the work.