Identification and Classification of Bus and Subway Passenger Travel Patterns in Beijing Using Transit Smart Card Data
Identification and Classification of Bus and Subway Passenger Travel Patterns in Beijing Using...
Wang, Lewen;Chen, Yuan;Wang, Yu;Sun, Xiaofei;Wu, Yizheng;Peng, Fei;Song, Guohua
Hindawi Journal of Advanced Transportation Volume 2023, Article ID 6529819, 15 pages https://doi.org/10.1155/2023/6529819 Research Article Identification and Classification of Bus and Subway Passenger Travel Patterns in Beijing Using Transit Smart Card Data 1 2 3 3 1 1 Lewen Wang , Yuan Chen, Yu Wang, Xiaofei Sun, Yizheng Wu , Fei Peng , and Guohua Song Key Laboratory of Transport Industry of Big Data Application Technologies for Comprehensive Transport, Beijing Jiaotong University, Haidian District, Beijing 100044, China Shanghai University of International Business and Economics, Shanghai, China CCCC Highway Consultants Co., Ltd., Dongcheng District, Beijing 100088, China Correspondence should be addressed to Yizheng Wu; firstname.lastname@example.org Received 6 September 2022; Revised 23 December 2022; Accepted 26 December 2022; Published 5 January 2023 Academic Editor: Domokos Eszterga´r-Kiss Copyright © 2023 Lewen Wang et al. Tis is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Urban public transit has been rapidly developed in recent years. However, given increases in travel volume, the level of service still needs to be improved to meet the satisfaction of passengers. Transit service providers and researchers have focused on improving transit devices, but the service level of public transit has not yet been efectively improved, so more and more research is interested in analyzing the travel patterns of passengers. Compared with traditional survey methods, smart card collection systems—which can collect spatial-temporal information about passengers’ trips—are convenient for the study of bus and subway passengers’ travel patterns. However, the data provided by smart cards have not yet been fully explored. Terefore, this paper proposed a multistep methodology to gather information on the travel patterns of bus and subway passengers in Beijing, China. We conducted statistical analyses and used an unsupervised clustering method to study and classify passengers based on travel patterns. Four groups have been identifed: standard commuters, fexible commuters, and two types of low-frequency passengers. Ten, a comprehensive analysis was conducted. We also discussed the changes of passengers’ travel time consumption before and after the implementation of customized bus for high-frequency passengers. Te analyses indicated that passengers’ travel patterns can provide useful information for transit service providers and can help improve the level of service of urban public transit by promoting the promulgation of local public transport policies and the implementation of customized services. problems have gradually emerged. According to the Beijing 1. Introduction Transport Annual Report, the total passenger fow in the Given increases in trafc congestion, the use of public transit public transit network decreased by 13.69% between 2014 is becoming more and more popular in many cities. Te and 2018, with a precipitous decline of almost 38% in 2020 development of urban public transit infrastructure has also compared to 2018 because of COVID-19. A similar trend was observed in the United States, suggesting that public attracted more people to use public transit, which may help to reduce pollutant emissions from motor vehicles , ac- transit has become less attractive and is facing greater celerate population mobility, alleviate trafc congestion, and challenges . To address declines in bus passenger fow, improve residents’ living standards. With the advantages of Nishiuchi et al.  analyzed temporal and spatial changes in convenience, afordability, and accessibility, more and more passengers’ travel behaviors in a month and revealed that people are willing to choose public transit . there are diferences in the travel patterns of diferent However, with the increased use of public transit and the passengers. Tey pointed out that more attention should be demand for personalized and diversifed travel, the public paid to the travel patterns of passengers. Transit agencies transit service is facing more challenges and several have been studying ridership patterns to identify key travel 2 Journal of Advanced Transportation factors such as ridership preferences and spatial and tem- of passengers’ travel patterns and clustered metro passengers poral patterns . By better understanding ridership pat- into four groups using statistical and unsupervised cluster- based methods. Yang et al.  classifed passengers into four terns, transit agencies can identify gaps in their current services and better adjust their operating strategies and types according to the number of routes and transfers and propose more relevant customized policies to improve proposed a method to estimate passenger spatial-temporal public transit services and encourage more usage. trajectory. To better understand long-term patterns in pas- Te advent of automated fare collection (AFC) systems senger travel, Kaewkluengklom et al.  studied changes in has provided great convenience, allowing for accurate and the individual travel behavior by using three years of longi- extensive recordings of each passenger’s travel transactions. tudinal smart card data from Shizuoka Prefecture, Japan, and Such data can be used for the extraction of passengers’ classifed passengers using the K-means method. Other boarding station , identifcation of commuters , and clustering methods have also been used in passenger classi- fcation. Kieu et al.  adopted the DBSCAN algorithm, analysis of passengers’ spatial-temporal dynamics [8, 9], among other possibilities. Terefore, to better explore the which can identify clusters of arbitrary shapes based on diferent parameters, to segment transit users into four travel regularity of passengers, identify their travel prefer- ences, and spatiotemporality patterns for proposing oper- groups. Tey stressed that passenger segmentation helps ational measures, we conducted a study using passenger operators to provide customized information and services for transaction data, established a method for dividing high- and diferent classes of transit users. Ouyang et al.  proposed a low-frequency passengers, identifed commuters, analyzed trip reconstruction algorithm and mined the travel patterns of the patterns of diferent group of passengers, proposed smart card users in Beijing using the DBSCAN algorithm. suggested measures, and briefy analyzed customized bus Also, Cui et al.  proposed a method to classify users based before and after implementation. on weekly boarding frequency using smart card data from Shenzhen collected over four consecutive weeks. Briand et al.  proposed a two-level model for passenger classifcation in 2. Literature Review Gatineau, Canada. Using Singapore as a model, Zhu  developed an activity type classifcation model by combining AFC systems are a valuable resource for studying passenger travel patterns in large cities. In recent years, using smart smart card data and traditional travel surveys to better un- derstand urban travel demand and activity dynamics. card data to analyze travel patterns has become a common technique employed by researchers. Despite the valuable insights provided by previous Since the frequency of passenger usage can afect the studies, the potential for smart card data to elucidate the development and proftability of urban public transit, many travel patterns of passengers has yet to be fully explored. studies have focused on identifying high-frequency pas- Firstly, identifying high-frequency passengers is usually sengers for better understanding their travel patterns to based only on the travel days or the trips, which is often not improve the services of public transit since they are the main comprehensive enough. In addition, the literature is less contributors. Te identifcation of high- and low-frequency likely to analyze the travel patterns of low-frequency pas- riders is usually based on researchers’ defnition to make an sengers after identifying high-frequency passengers yet artifcial distinction. Nishiuchi et al.  classifed the top improving the utilization rate of low-frequency passengers is 40% as low-frequency passengers based on the cumulative one of the keys to improving service quality. curve of trip days, who traveled only 1 or 2 days during the To build on previous studies, using smart card data to period, while the remaining 60% were high-frequency examine passenger patterns, this paper proposed a multistep passengers. Kieu et al.  identifed frequent passengers by methodology to mine the travel patterns of bus and subway having 53 trips on weekdays for three months, i.e., at least passengers in Beijing, China. We studied passenger travel one trip per day during at least 75% of the weekdays. Lathia patterns by analyzing three main variables: trip frequency, et al.  and El Mahrsi et al.  used 2 out of 30 days and travel time, and travel space. Te approach used in this study 10 out of 30 days as frequency thresholds to distinguish high- is shown in Figure 1. First, we used smart card data and frequency passengers from low-frequency passengers, re- geographic information data to extract each passengers’ spectively. Liu and Cheng  defned high-frequency travel indicators. Ten, combined with the statistical analysis passengers as those who used their Oyster cards at least method, we analyzed passengers’ general travel patterns. In 37.5% of days during the four-week study period. While addition to using the travel days and trips, we established a Zhao et al.  defned commuters as the 95th percentile method to identify high- and low-frequency passengers by ratio of routine weeks among all smart card holders. combining weekday and week-based travel frequency. We In addition to trip frequency, the temporal and spatial also used a clustering-based analysis method to group indicators of passenger are often used for passenger classi- passengers based on their multidimensional travel indica- fcation and combined with clustering algorithms. Te tors, then mined their travel patterns in detail. Tis paper K-means algorithm is a frequently used algorithm with simple makes the following contributions: parameters, as the only required input is the number of (i) We extracted the trip frequency index, temporal clusters (K). Ortega-Tong  divided London public transport users into four categories based on travel frequency, index, and spatial index for each passenger and analyzed the general travel patterns of passengers travel hour, travel time consumption, and travel start and end stations. Zhao et al.  analyzed the spatial-temporal aspects based on these three indicators. Journal of Advanced Transportation 3 Smart card data + Geographic information data Travel pattern extraction + Clustering-based analysis Statistical-based analysis General travel patterns of all passengers Travel patterns of different groups of passengers Figure 1: Overview of the approach used in this study to examine passenger travel patterns in Beijing. (ii) We established a method for identifying high- and Transport Survey Summary Report in Beijing ), meaning low-frequency passengers considering multiple trip that if the time between the end time of the last trip and the frequency indicators and classifed passengers into start time of the current trip was less than 30 minutes, the diferent groups using statistical methods and un- two subtrips were considered as one trip. Also, the data felds supervised clustering methods. and their defnitions are shown in Table 1. (iii) We conducted a comprehensive analysis of both high-frequency and low-frequency passenger pat- 3.3. Travel Index Extraction and Analysis terns based on trip frequency, travel time, travel space, transfers, and made targeted recommenda- 3.3.1. Trip Frequency Index Extraction. We extracted fve tions and analyzed the efects of the measures before indicators for each passenger: monthly travel days (D ), and after implementation, which can provide a basis monthly trips (T ), monthly travel days on working days for improving urban public transit services. (D ), monthly trips on working days (T ), and number of mw mw weeks with more than three travel days (C ). 3. Experimental Methods and Materials Te distribution of travel days and trips showed a bi- modal pattern (Figure 3). Te majority of passengers with 3.1. Study Area. Te smart card dataset used in this study fewer travel days and trips, but there still existed some was collected from Beijing, the capital of China, which is passengers who traveled very frequently. It is important to composed of 16 administrative regions. Beijing has a resi- capture the travel patterns of these passengers, which will be dent population of nearly 21.886 million, and the city explained in later sections. contains about 1,200 bus lines and 27 subway lines. Due to the large population and the large daily demand on public transit in Beijing, the number of daily card swipes can reach 3.3.2. Temporal Index Extraction. As passengers usually the tens of millions. Tus, we restricted our data collection follow a fxed pattern on weekdays, we extracted the average and analysis to two large residential areas: Tiantongyuan and departure hours for the frst departure (H ) and the mw Huilongguan, which are both located in the Changping variance of the frst departure time interval (V ) on mw District (Figure 2). Te resulting dataset included more than weekdays. 1.8 million travel records of 70,539 passengers from Sep- From Figure 4, the total amount of trips spiked during tember 2019. peak hours and remained stable during nonpeak hours, which showed a doubled peak on each day, but more 3.2. Dataset. Te data used in this study were preliminarily pronounced on weekdays. As passengers prefer to travel on processed, including station matching and transfer identi- weekdays, the trips on weekdays were higher than weekends. fcation. A trip may include multiple subtrips, which is commonly referred to as a transfer; transfers may occur between the same modes of transportation (e.g., bus to bus; 3.3.3. Spatial Index Extraction. We mainly considered the subway to subway cannot be recorded because passengers working days with closed travel chains (TC ) and the mw only swipe their cards in and out of the subway station) or average travel distance (Dis ). Supposing that the origin (O) between diferent modes (e.g., subway to bus or bus to of the ith trip by public transit on day d is O , the destination di subway). We set the transfer time threshold as 30 minutes (D) is D , and the total number of trips in the day is i , the di max (this threshold was set based on the 5th Comprehensive set of O-D pairs on day d can be expressed as follows: 4 Journal of Advanced Transportation Dongcheng 41.0 Shijingshan Huairou 40.8 Xicheng 40.6 Yanqing Miyun 40.4 Changping Pinggu 40.2 Shunyi Haidian 40.0 Mentougou Chaoyang Fengtai 39.8 Fangshan Tongzhou Daxing 39.6 39.4 10 km 115.5 116.0 116.5 117.0 117.5 Longitude Tiantongyuan Huilongguan Figure 2: Study area. Table 1: Smart card data format. Field Defnition CNID Card number. Each smart card has a diferent card number. Trip Itinerary number. Each trip has a diferent itinerary number. Subtrip number. Number of increments in each trip: if the trip does not include a transfer, the column will be 0; if it includes SubTrip transfer, it will increase from 1. TrTp Te mode of transportation used in the trip. 1 � bus only. 2 � subway only. 3 � bus and subway. SubtrTP Te type of transportation mode used in the subtrip. b � bus, s � subway. T1 Te time that passengers swipe their cards at the boarding station. STATION1 Te name of the boarding station. lon1 Te longitude of the boarding station. lat1 Te latitude of the boarding station. T2 Te time that passengers swipe their cards at the alighting station. STATION2 Te name of the alighting station. lon2 Te longitude of the alighting station. lat2 Te latitude of the alighting station. A ⟶B and A ⟶B , despite having diferent starting and 1 1 2 2 O − D , O − D , O − D , . . . , O − D . (1) d1 d1 d2 d2 d3 d3 di di max max ending stations, are both O⟶D trips. Tus, in our study, if As passengers often go back and forth between O and D, the distance between the frst departure station (O ) and the d1 it may be through A and B of line 1 or A and B of line 2, last arrival station (D ) in a day is less than 1.2 km , it 1 1 2 2 dimax where A and A are close to O and B and B are close to D; is considered to constitute a closed travel chain (more details 1 2 1 2 the number of lines will vary depending on the case. can be found in the Supplementary Materials). In examining Latitude Journal of Advanced Transportation 5 0.08 0.175 0.07 0.150 0.06 0.125 0.05 0.100 0.04 0.075 0.03 0.050 0.02 0.025 0.01 0.000 0.00 0 10 20 30 0 50 100 150 200 Monthly travel days Monthly trips 0.10 0.25 0.08 0.20 0.06 0.15 0.04 0.10 0.02 0.05 0.00 0.00 05 10 15 20 0 50 100 150 Monthly travel days on working days Monthly trips on working days Figure 3: Trip frequency of passengers in a month, including monthly travel days, monthly trips, monthly travel days on working days, and monthly trips on working days. characteristics, with passengers moving from outside to average travel distance, we used the average travel time cost as a proxy for travel distance as has been done in previous inside the city in the morning peak, and vice versa in the evening peak. Tis was because many passengers were studies . Most passengers’ travel time cost were about 50 minutes commuters who return to their residence after completing (Figure 5(a)). Tere was an approximately linear relationship their trips during the day. Tis was also indicated the between the number of transfers and the average travel time separation of workplace and residences that existed in cost (Figure 5(b)), which suggested that transfers have an Beijing, consistent with the study of Ma et al. . important impact on passengers’ travel time cost. Tere were diferences in the distribution of passenger 4. Passenger Classification boarding and alighting stations during the morning and evening peak periods. From Figure 6(a), it can be found that 4.1. Passenger Classifcation Based on Trip Frequency. passenger boarding stations during the morning peak were From the previous analysis, we knew that passengers vary mainly concentrated in large residential area- widely in their trip frequency. Tus, to better analyze and s—Tiantongyuan and Huilongguan areas. Figure 6(b) shows highlight the regularity of habitual passengers, we proposed that alighting stations were mainly distributed in the central a set of methods considering multidimensional travel fre- city, where many enterprises and schools are gathered. quency indicators. In this paper, we considered high-fre- Due to the daytime trips, passengers have spread all over quency passengers to those whose (a) monthly travel days the city. Terefore, from Figure 6(c), boarding stations in the D ≥ 14 and monthly travel trips T ≥ 18, or (b) travel days m m evening peak were relatively dispersed. However, the on weekdays D ≥ 10 and trips on weekdays T ≥ 15, or mw mw alighting stations were mainly concentrated in large resi- (c) under the condition of D ≥ 10 and T ≥ 15, the number m m dential areas (Figure 6(d)). Te movement of passengers in of weeks with more than three travel days C ≥ 3. After a the morning and evening peaks showed diferent preliminary analysis (the detail can be seen in the Density Density Density Density 6 Journal of Advanced Transportation 8000 8000 6000 6000 4000 4000 2000 2000 2000 0 0 0 0 612 18 24 0 6 12 18 24 0 6 12 18 24 Mon. Tue. Wed. 2000 1000 0 0 0 0 6 12 18 24 0 6 12 18 24 0 6 12 18 24 Time of day Time of day Time of day Tu. Fri. Sat. Sun. Figure 4: Total amount of trips per day of a week. Supplementary Materials), 23,103 people met condition (a), passenger and any other cluster. Te value of the average 23,948 people met condition (b), and 24,530 people met silhouette coefcient is contained within (− 1, 1); values condition (c). Ultimately, 26,100 high-frequency passengers closer to 1 indicate a better clustering efect. and 44,439 low-frequency passengers were identifed. By clustering high-frequency passengers and low-fre- Te classifcation results are shown in Table 2, and high- quency passengers separately, we found that the clustering frequency riders traveled an average of 20.09 days and 40.51 efect was optimal when the indicators (H , V , TC , mw mw mw trips per month, while low-frequency riders traveled and Dis ) and (D , T , D , T , H , V , TC , and m m m mw mw mw mw mw 4.31 days and 7.31 trips per month. Although the percentage Dis ) were selected for the high- and low-frequency pas- of high-frequency passengers (37%) was lower than low- senger groups, respectively, and the numbers of clusters frequency passengers (63%), they accounted for 76.5% of all were 2. passenger trips. 5. Analysis of Different Groups of Passengers 4.2. Passenger Clustering Based on Multiple Indicators. 5.1. Analysis of Travel Frequency. Figure 7 shows that low- Since the K-means algorithm only needs to input the frequency passengers traveled less frequently than high- classifcations K to get the classifcation labels of all pas- frequency passengers. Te frst group of high-frequency sengers, it is very efcient and suitable for the huge data set passengers traveled most frequently; they usually traveled on of urban public transit. And the passenger classifcation working days and their trips were twice as frequent as travel method using K-means algorithm has been widely used in days, which meant they often traveled twice a day, likely combination with some evaluation methods [9, 17]. Tus, we between residence and workplace, and they were usually used the K-means algorithm combined with the average commuters. As for transfers, nontransfer trips accounted for silhouette coefcients to select the best combination of in- the majority (Figure 8). Tis suggested that most passengers dicators and the optimal number of clusters from the set of prefer to travel without transferring, indicating urban public indicators (D , T , D , T , H , V , TC , and Dis ). m m mw mw mw mw mw m operators should design their routes with the goal of min- Te average silhouette coefcient is calculated as follows: imizing transfers. a − b i i s � , (2) max a , b i i 5.2. Temporal Analysis. Te frst departure hour for two where a is the average dissimilarity between the ith pas- group high-frequency passengers was concentrated in the senger and all other passengers within the same cluster and morning peak. However, the wide distribution for two low- b is the lowest average dissimilarity between the ith frequency groups suggested that they were more fexible in Total amount of trips Total amount of trips Journal of Advanced Transportation 7 0 50 100 150 200 Average time cost (minutes) (a) Transfer Regression line 95% Confdence interval (b) Figure 5: Travel time cost pattern of all passengers. (a) Te average travel time cost for all passengers. (b) Relationship between average travel time cost and transfer. their travel times (Figure 9). Figure 10 shows high-frequency 5.3. Spatial Analysis. Average travel time cost was similar passengers exhibit a clear bimodal travel pattern on both among the diferent types of passengers (median around weekdays and nonweekdays. In contrast, low-frequency 50 minutes and similar shape of distribution), as they tended to pursue the shortest travel time (by reducing transfers, passengers did not show a clear pattern and their travel times were relatively scattered. In addition, trips made by high- Figure 12). Te frst type of high-frequency passengers had frequency passengers on weekends decreased to about half of the most closed travel chains on weekdays, which meant they on weekdays, while low-frequency riders remained constant. were more stable in travel space (Figure 13), and more According to Figure 11, the median of the frst group of analysis can be seen in the Supplementary Materials. As high-frequency passengers was the lowest and showed a analyzed above, these group passengers also stable in the single-peaked distribution, mostly distributed around 0, temporality dimension. Terefore, the frst group of high- implying that this group started their travel plans at a fxed frequency passengers were mostly standard commuters, who time each weekday. Te second group of high-frequency traveled to and from the fxed area (residence and work- passengers had a more even distribution, but also had a place) every day during peak hours. Te second type of high- higher percentage of passengers with lower travel time frequency passengers were fexible commuters and their last variances. In addition, the median of the frst low-frequency trip was freer. We used OD diagrams to visualize the travel passengers was lower, indicating that they maintained sta- records of typical passengers in each group (Figure 14). Te bility in travel time for lower frequency trips. Te second OD of high-frequency passengers were more clustered into a group showed a uniform distribution. However, there were circle, especially for the frst type of high-frequency pas- passengers with large travel time interval variance among sengers, while the OD of low-frequency passengers were low-frequency passengers. more scattered, which forms multiple circles. Average time cost (minutes) Passengers 8 Journal of Advanced Transportation (a) (b) (c) (d) Figure 6: Heat maps for boarding and alighting stations in the morning and evening peaks. (a) Boarding stations in the morning peak. (b) Alighting stations in the morning peak. (c) Boarding stations in the evening peak. (d) Alighting stations in the evening peak. Table 2: Classifcation results for high-frequency and low-frequency passengers. Monthly trips Monthly travel days Group Passenger proportion (%) Trip proportion (%) Average Standard deviation Average Standard deviation Low-frequency 63 23.50 7.31 14.75 4.31 4.75 High-frequency 37 76.50 40.51 5.89 20.09 3.05 Point D in Figure 15 was Tiantongyuan and Hui- frequency passengers have diferent patterns than low-fre- longguan area in the Changping District, which is a large quency passengers. Tey showed more travel activity in the residential area in Beijing. We inferred that most of the center of the city. For example, they often traveled between passengers in the dataset lived in this area. From Tiantongyuan and Huilongguan areas and Xierqi Figure 15(a), the frst group of low-frequency passengers (Figure 15(b)) where there was a Zhongguancun Software Park with a lot of ofce buildings and enterprises. And the mainly traveled between Tiantongyuan and Huilongguan (point D) and Beijing West Station (point A), Beijing South bar chart showed the top 10 OD pair trips among 10,473 and Station (point B), or Beijing Station (point C), and bar chart 21,158 OD pairs for two group of high-frequency passengers, showed the top 10 OD pair trips among 3,543 OD pairs. Te respectively. Tis also showed that high-frequency riders second group showed a similar pattern, but it was not as travel between residence and workplace in high volumes and pronounced as the frst group, with the bar chart showing travel over a wider range of travel spaces than low-frequency the top 10 OD pair trips among 4532 OD pairs. High- riders. Journal of Advanced Transportation 9 0123 0123 Legend 20.0 150 Outside Points Upper Adjacent Value Tird Quartile 17.5 Median 15.0 First Quartile Lower Adjacent Value Group 1 of low-frequency 12.5 Group 2 of low-frequency Group 1 of high-frequency Group 2 of high-frequency 10.0 7.5 5.0 2.5 0123 0123 Group of passengers Group of passengers Figure 7: Travel frequency for four types of passengers. No transfer No transfer 71.00% 71.00% 29.00% 29.00% Transfer Transfer Group 1 of low-frequency passengers Group 2 of low-frequency passengers No transfer No transfer 66.00% 81.00% 19.00% 34.00% Transfer Transfer Group 1 of high-frequency passengers Group 2 of high-frequency passengers Figure 8: Te proportion of diferent groups of passengers with or without transfer trips. Travel days on working days Monthly travel days Trips on working days Monthly trips 10 Journal of Advanced Transportation 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Time of day Group 1 of low-frequency passengers Group 2 of low-frequency passengers Group 1 of high-frequency passengers Group 2 of high-frequency passengers Figure 9: Average departure hours for four types of passengers. Working day Nonworking day 0369 12 15 18 21 24 (0) 369 12 15 18 21 Time of day Group 1 of low-frequency passengers Group 2 of low-frequency passengers Group 1 of high-frequency passengers Group 2 of high-frequency passengers Figure 10: Comparison of travel hours between working days and nonworking days for diferent groups of passengers. Te spatial pattern of high-frequency commuters 6. Discussion showed that they travel more between large residential Te purpose of studying passenger travel characteristics is to areas and typical workplaces, so customized buses can be identify passengers’ travel patterns of diferent groups of considered, and the departure time and frequency can be passengers. Tis is helpful to propose suggestions for im- set based on the identifed travel patterns of commuters. provement, formulate corresponding fare and service pol- Figure 15(b) shows that some commuters travel fre- icies, and make urban public transit more attractive. quently between Tiantongyuan and Xierqi, so we used the Terefore, based on the results of the above analysis, we example of setting up the customized bus from Tian- illustrated how the ridership pattern analysis provided the tongyuan to Xierqi for the discussion of individualized basis for the relevant customized services. measures. Trips Passengers Journal of Advanced Transportation 11 Legend Outside Points Upper Adjacent Value Tird Quartile Median First Quartile Lower Adjacent Value Group 1 of low-frequency Group 2 of low-frequency Group 1 of high-frequency Group 2 of high-frequency Group of passengers Figure 11: Variance in the frst departure time interval for various passengers. Legend Outside Points Upper Adjacent Value Tird Quartile Median First Quartile Lower Adjacent Value Group 1 of low-frequency Group 2 of low-frequency Group 1 of high-frequency Group 2 of high-frequency Group of passengers Figure 12: Distribution of average travel time for four groups of passengers. Legend Outside Points Upper Adjacent Value Tird Quartile Median First Quartile Lower Adjacent Value Group 1 of low-frequency Group 2 of low-frequency Group 1 of high-frequency Group 2 of high-frequency 012 3 Group of passengers Figure 13: Closed trip chain distribution for four groups of passengers. Figure 16(a) shows the average travel time con- Maps, the minimum time cost from Tiantongyuan to sumption of high-frequency passengers who traveled Xi’erqi in the morning peak is 30 minutes. Terefore, we between Tiantongyuan and Xierqi. Most passengers assumed that the travel time consumption of the cus- traveling in the 30–40 minute range, with an average travel tomized bus is 30 minutes. Figure 16(b) shows the change time cost of 39.21 minutes. Customized bus is designed of time cost per passenger after choosing customized bus, primarily for buses to provide a higher level of service, with the majority of passengers experiencing a signifcant faster and more comfortable for passengers . reduction, for a total reduction of 2,550 minutes for all According to the real-time prediction function of Google passengers. Working days with closed travel chain Time interval variance Travel time cost (minutes) 12 Journal of Advanced Transportation Group1 of low-frequency passengers Group2 of low-frequency passengers 40.15 40.15 Shunyi Shunyi Changping Changping 40.10 40.10 40.05 40.05 Haidian Haidian 40.00 40.00 Chaoyang 39.95 Chaoyang 39.95 Xicheng Xicheng 39.90 39.90 Dongcheng Dongcheng 39.85 39.85 Fengtai Fengtai 39.80 39.80 Tongzhou Tongzhou 39.75 39.75 Daxing Daxing 39.70 39.70 116.2 116.3 116.4 116.5 116.6 116.7 116.2 116.3 116.4 116.5 116.6 116.7 Longitude Longitude Group1 of high-frequency passengers Group2 of high-frequency passengers 40.15 40.15 Shunyi Shunyi Changping Changping 40.10 40.10 40.05 40.05 Haidian Haidian 40.00 40.00 Chaoyang 39.95 39.95 Chaoyang Xicheng Dongcheng 39.90 39.90 Dongcheng Xicheng 39.85 39.85 Fengtai Fengtai 39.80 39.80 Tongzhou Tongzhou 39.75 39.75 Daxing Daxing 39.70 39.70 116.2 116.3 116.4 116.5 116.6 116.7 116.2 116.3 116.4 116.5 116.6 116.7 Longitude Longitude Figure 14: OD diagram for a typical passenger selected from each group in a month. Trips D: Tiantongyuan and Huilongguan area D→A 20.0 D→C A→D A: Beijing West Station B: Beijing South Station C→D 17.5 C: Beijing Station D→B 15.0 B→D 12.5 10.0 7.5 5.0 2.5 0.0 01234 56789 OD pair Trips D→B D→C B→D D→A 01234 56789 OD pair (a) Figure 15: Continued. Latitude Latitude Group 2 of low-frequency passengers Group 1 of low-frequency passengers Journal of Advanced Transportation 13 Trips D: Tiantongyuan and Huilongguan area E→D D→E E: Xierqi 01234 56789 OD pair Trips D→E E→D 01234 56789 OD pair (b) Figure 15: Travel spatial patterns for four types of passengers on weekdays. (a) Low-frequency passengers. (b) High-frequency passengers. Average time cost of a trip (minutes) (a) Figure 16: Continued. Percent (%) Group 2 of high-frequency passengers Group 1 of high-frequency passengers Cumulative percentage (%) 14 Journal of Advanced Transportation Increase –10 –20 –30 –40 –50 0 50 100 150 200 250 Passenger No. Change in time cost (b) Figure 16: Te travel time cost of high-frequency passengers with trips from Tiantongyuan to Xierqi. (a) Average travel time cost per trip for passengers using public transit. (b) Change in passenger travel time cost after selecting customized bus. In addition, high-frequency passengers travel much more combine other data for passenger profling to analyze pas- frequently than low-frequency passengers, so discounted fares senger characteristics from a multidimensional perspective and souvenirs can be ofered to passengers who meet the high- [27, 28]. frequency ridership classifcation, which can also motivate more low-frequency passengers who prefer to choose other Data Availability travel modes (e.g., owned vehicles) to use public transit. Also, low-frequency passengers can be encouraged by reducing fares Te passenger transactions used to support the fndings of at nonpeak hours as they chose their travel time more freely. this study are available from the corresponding author upon request. 7. Conclusions Conflicts of Interest In this paper, we studied the travel patterns of public transit passengers in Beijing, including both general patterns and Te authors declare that there are no conficts of interest the diferent patterns for four groups of passengers. First, we regarding the publication of this paper. extracted travel indicators for each passenger and conducted a general analysis. Ten, we used an unsupervised clustering Acknowledgments algorithm to classify passengers and analyzed the travel frequency, travel time, travel space, transfers, and travel time Tis work was supported by the Fundamental Research cost for the four resulting groups of passengers. We found Funds for the Central Universities (grant no. 2022JBMC056) that there were some passengers with high travel frequency and the National Natural Science Foundation of China whose travel exhibited stable temporal and spatial patterns. (grant no. 71901018 and 52272340). All the data used for this Most of these passengers were commuters (standard com- study were provided by the Beijing Transport Institute. muters and fexible commuters), which comprise the main users of urban public transit. And it is feasible to improve Supplementary Materials public transit service levels based on passenger travel pat- terns. In addition, there are many limitations in our study. Te supplementary material consisted of three sections. Te For example, we only studied the travel frequency and travel frst was an explanation of the choice of 1.2 km in the closed spatial and temporal characteristics of passengers, but not travel chain. Te second was a methodological supplement other aspects of travel characteristics. In the future, we will to the classifcation of high- and low-frequency passengers. Decrease or increase in passenger time cost afer selecting DRT (minutes) Journal of Advanced Transportation 15  X. Zhao, M. Cui, and D. Levinson, “Exploring temporal Te third was a supplement to the spatiotemporal patterns of variability in travel patterns on public transit using big smart diferent groups of passengers. (Supplementary Materials) card data,” Environ Plan B Urban Anal City Sci, SAGE Publications Sage UK, London, England, Article ID References 23998083221089664, 2022.  M. A. Ortega-Tong, Classifcation of London’s Public Trans-  M. Chester, S. Pincetl, Z. Elizabeth, W. Eisenstein, and port Users Using Smart Card Data, Massachusetts Institute of J. Matute, “Infrastructure and automobile shifts: positioning Technology, Cambridge, MA 02139, USA, 2013. transit to reduce life-cycle environmental impacts for urban  T. Yang, P. Zhao, and X. Yao, “A method to estimate URT sustainability goals,” Environmental Research Letters, vol. 8, passenger spatial-temporal trajectory with smart card data no. 1, Article ID 15041, 2013. and train schedules,” Sustainability, vol. 12, no. 6, p. 2574,  Y. Sun and Y. Cui, “Evaluating the coordinated development of economic, social and environmental benefts of urban  R. Kaewkluengklom, F. Kurauchi, and T. Iwamoto, “Inves- public transportation infrastructure: case study of four Chi- tigation of changes in passenger behavior using longitudinal nese autonomous municipalities,” Transport Policy, vol. 66, smart card data,” International Journal of Intelligent Trans- pp. 116–126, 2018. portation Systems Research, vol. 19, no. 1, pp. 155–166, 2021.  M. E. G. Parker, M. Li, M. A. Bouzaghrane et al., “Public  L. M. Kieu, A. Bhaskar, and E. Chung, “Transit passenger transit use in the United States in the era of COVID-19: transit segmentation using travel regularity mined from Smart Card riders’ travel behavior in the COVID-19 impact and recovery transactions data,” Transportation Research Board (TRB) 93rd period,” Transport Policy, vol. 111, pp. 53–62, 2021. Annual Meeting Compendium of Papers, pp. 1–17, Wash-  H. Nishiuchi, J. King, and T. Todoroki, “Spatial-temporal ington, D.C, USA, 2014. daily frequent trip pattern of public transport passengers  Q. Ouyang, Y. Lv, Y. Ren, J. Ma, and J. Li, “Passenger travel using smart card data,” International Journal of Intelligent regularity analysis based on a large scale smart card data,” Transportation Systems Research, vol. 11, no. 1, pp. 1–10, 2013. Journal of Advanced Transportation, vol. 2018, Article ID  X. Ma, Y. J. Wu, Y. Wang, F. Chen, and J. Liu, “Mining smart 9457486, 2018. card data for transit riders’ travel patterns,” Transportation  C. Cui, Y. Zhao, and Z. Duan, “Research on the stability of Research Part C: Emerging Technologies, vol. 36, pp. 1–12, public transit passenger travel behavior based on smart card data,” in Proceedings of the CICTP 2014: Safe, Smart, and  X. Ma, Y. Wang, F. Chen, and J. Liu, “Transit smart card data Sustainable Multimodal Transportation Systems, pp. 1318– mining for passenger origin information extraction,” Journal 1326, Changsha , China, July 2014. of Zhejiang University - Science C, vol. 13, no. 10, pp. 750–760,  A. S. Briand, E. Come, ˆ M. Tr´epanier, and L. Oukhellou, “Analyzing year-to-year changes in public transport pas-  C. M. Schneider, C. Rudlof, D. Bauer, and M. C. Gonz alez, ´ senger behaviour using smart card data,” Transportation “Daily travel behavior: lessons from a week-long survey for Research Part C: Emerging Technologies, vol. 79, pp. 274–289, the extraction of human mobility motifs related information,” in Proceedings of the 2nd ACM SIGKDD international  Y. Zhu, “Estimating the activity types of transit travelers using workshop on urban computing, pp. 1–7, Chicago, Illinois, smart card transaction data: a case study of Singapore,” August 2013. Transportation, vol. 47, no. 6, pp. 2703–2730, 2020.  S. Tao, D. Rohde, and J. Corcoran, “Examining the spatial--  Y. Zhang, Z. Zhou, W. Wang, and R. Zhang, Te 5th Com- temporal dynamics of bus passenger travel behaviour using prehensive Transport Survey Summary Report, Beijing China, smart card data and the fow-comap,” Journal of Transport Geography, vol. 41, pp. 21–36, 2014.  Y. Zhao, G. Yin, and S. Wu, 2021 China Principal Cities  J. Zhao, Q. Qu, F. Zhang, C. Xu, and S. Liu, “Spatio-temporal Sharing Bikes and Sharing Electric Bikes Riding Report, Beijing analysis of passenger travel patterns in massive smart card China, 2021. data,” IEEE Transactions on Intelligent Transportation Sys-  X. Ma, C. Liu, H. Wen, Y. Wang, and Y. J. Wu, “Under- tems, vol. 18, no. 11, pp. 3135–3146, 2017. standing commuting patterns using transit smart card data,”  L. M. Kieu, A. Bhaskar, and E. Chung, “Mining temporal and Journal of Transport Geography, Pergamon, vol. 58, pp. 135– spatial travel regularity for transit planning,” in Proceedings of 145, 2017. the Australasian Transport Research Forum 2013 Proceedings,  A. Huang, Z. Dou, L. Qi, and L. Wang, “Flexible route op- pp. 1–12, Australasian Transport Research Forum, Brisbane, timization for demand-responsive public transit service,” QLD Australia, October 2013. J Transp Eng A Syst, vol. 146, no. 12, Article ID 04020132, 2020.  N. Lathia, D. Quercia, and J. Crowcroft, “Te hidden image of the city: sensing community well-being from urban mobility,” in Proceedings of the International Conference on Pervasive Computing, pp. 91–98, Springer, Berlin, Heidelberg, June  M. K. El Mahrsi, E. Coˆme, J Baro, and L Oukhellou, “Under- standing passenger patterns in public transit through smart card and socioeconomic data: A case study in rennes, france,” ACM SIGKDD workshop on urban computing, New York, NY, USA, 2014, https://www.comeetie.fr/pdfrepos/urbcomp2014.pdf.  Y. Liu and T. Cheng, “Understanding public transit patterns with open geodemographics to facilitate public transport planning,” Transportmetrica: Transportation Science, Taylor & Francis, vol. 16, no. 1, pp. 76–103, 2020.
Journal of Advanced Transportation
Hindawi Publishing Corporation