1. Introduction
In the context of climate change and rapid urbanization, extreme rainfall events tend to increase in intensity and frequency (Sillmann et al. 2013; Pendergrass and Hartmann 2014; IPCC 2021). For example, observations showed a significant upward trend in maximum hourly rainfall in Shanghai, China, from 1916 to 2014, with increases of 2.72 and 6.2 mm h−1 (10 yr)−1 in the first and second half of the period, respectively (Liang and Ding 2017). In the second half of the period, the average value of maximum hourly rainfall over the last decade (2005–14) reached 85.5 mm h−1, about 1.5 times as high in comparison with the 57 mm h−1 in the first decade (1965–74). During the rapid urbanization period (1981–2014), the annual frequency of heavy hourly rainfall in Shanghai increased significantly at a rate of 1.8 (10 yr)−1. Spatially, heavy rainfall events were increasingly concentrated in urban and suburban areas of the city (Liang and Ding 2017). The concurrence between rising rainfall events and traffic congestion often poses significant stresses on urban transport system (Yang et al. 2016). It is of great significance to quantitatively assess the impact of rainfall on urban commuting, and to improve the coping capacity and resilience of transport systems accordingly (Böcker et al. 2013; Kashfi et al. 2016; Diab and Shalaby 2020). The metro plays an important role in commuting in Shanghai (Su et al. 2016), accounting for nearly 60% of commuting flows in the central city in 2020. Exploring the spatiotemporal variations of metro commuting demand, behavior and flows during rainfall events is fundamental to quick and accurate spatial and temporal response decisions (Tang et al. 2018).
Questionnaires are commonly used to investigate the impact of rainfall on commuters’ demand and behavior, such as commuting intention, mode, departure time, and temporal distance. Zanni and Ryley (2015) found that commuters would rarely change their plans during inclement weather as compared with travelers for other purposes, because the importance of the commuting activity outweighed the negative impact of inclement weather. Wu and Liao (2020) pointed out that respondents preferred to commute by metro or car during rainfall. Cools and Creemers (2013) concluded that commuters tended to advance or delay their departure time during rainfall. Jain and Singh (2021) found that metro commuters with longer distance were less likely to change their regular travel mode during rainfall. However, there are often inconsistencies between the questionnaire-based findings and the actual travel actions that depend not only on travelers’ willingness to travel but also on the purpose of their travels and the surroundings where they are located.
Over the last decade, real-time passenger flow data have been increasingly used to explore the impact of extreme weather on different travel groups and purposes (Zhou et al. 2017; Wei et al. 2021; Zhou et al. 2021). Arana et al. (2014) studied the impact of extreme weather on different travel purposes and found that regular travelers were less affected than irregular ones. Tao et al. (2016) examined the impact of severe weather on commuting flows at different spatial scales using a deviation correlation model and geographical visualization techniques. Tao et al. (2018) applied the time series regression models to assess the changes in commuting flows resulting from weather factors. Most of these studies focus on the analysis of the overall relationship between daily rainfall and daily commuting flows based on long time series data (Kalkstein et al. 2009; Li et al. 2018).
Changes in commuting flows are influenced by a variety of factors including the onset time, duration, intensity and spatial extent of rainfall events, the distribution of metro stations and their surrounding built-up environment, and the departure time and travel distance of commuters. The daily scale studies aforementioned cannot reveal the spatiotemporal patterns of commuting flows in detail, nor can they identify the times and areas that are most sensitive to rainfall. In recently years, a few studies have focused on the relationship between ridership and weather condition at finer spatiotemporal scales. For example, Zhou et al. (2017) distinguished between peak and off-peak times and analyzed the variation of intraday metro ridership caused by weather conditions at both system and station levels. Najafabadi et al. (2019) used a Bayesian multilevel regression model to examine the impact of rainfall on hourly and daily metro ridership at individual stations. However, the above studies did not analyze the hourly ridership differences specifically in the peak hours in detail, nor did they consider the origin–destination (OD) flow across stations.
Therefore, based on hourly and spatial rainfall data, hourly OD flow data across stations, and built environment data, this study explores the variation of metro flows due to rainfall at a finer spatiotemporal scale. It focuses on hourly commuting flows in the peak hours, both at station and OD levels. The results are expected to provide a reasonable basis for making spatial and temporal response strategies.
2. Data and methods
a. Study area
The study area in Fig. 1 covers the Shanghai metropolitan area, including 15 metro lines and 320 metro stations. Note that the “metro” comprises the underground subway, light-rail line, and commuting railroad. A few newly opened lines are excluded because of data limitation.
Distribution of metro lines and stations in Shanghai.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
b. Data collection and processing
1) Rainfall data
Hourly rainfall data are crawled from the official website (http://bmxx.swj.sh.gov.cn/shfx) established by the Shanghai Water Authority (Shanghai Municipal Oceanic Bureau), consisting of the rainfall events in the three days of 5 June, 6 July, and 15 July 2020. The rainfall events in these three days coincide highly with the morning and evening peak hours and are therefore chosen as typical cases in this study.
The data are classified into four levels according to rainfall intensity, that is, light (≤2.5 mm h−1), moderate (2.6–8.0 mm h−1), heavy (8.1–15.0 mm h−1), and extremely heavy (≥16.0 mm h−1) (Wang 2018). Figure 2 shows the rainfall process in the three days and its spatial distribution using kriging interpolation.
(left) Hourly rainfall and (right) its spatial distribution in Shanghai on (a) 5 Jun 2020, (b) 6 Jul 2020, and (c) 15 Jul 2020.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
2) Metro OD flow data
The metro OD flow data are obtained from the Shanghai Urban–Rural Construction and Transportation Development Research Institute, spanning the periods from 6 May to 29 July 2020, excluding weekends. Note that only the metro OD flow data from the three days of 5 June, 6 July, and 15 July are used in the metro flow change analysis due to rainfall. The flow data from the other days during 6 May–29 July are used as the training data in the Prophet time series model (Taylor and Letham 2018) to predict the baseline flow in the above three days. To quantify the change in commuting flows due to rainfall, it is necessary to compare the observed flow in the case of rainfall with the baseline flow at the same time with no rainfall. This baseline flow cannot simply be taken as the observed data from a particular day or days in the past because the observed value is subject to nonperiodic trends, periodic variations, holidays or events, and random errors. Therefore, it often needs to be predicted using time series models, such as the Prophet model. The Prophet model will be described in detail in section 2c.
The data are structured as hourly OD pairs (0600–2200 LT), containing the inbound station code, outbound station code and passenger flows (Table 1).
Example of data records for metro commuting flows in Shanghai.
In addition, the OD flow data are applied to the metro station distribution map in Fig. 1, which are then overlaid with the rainfall distribution map in Fig. 2. This allows the hourly passenger flows to be spatially correlated with rainfall at individual stations. Based on the above integrated multisource data with high temporal and spatial resolution, the impact of rainfall on metro passenger flows can be explored in detail.
c. Methods
1) Research framework
This study is intended to analyze the impact of rainfall on commuting flows. Therefore, we focus on the flows during commuting hours, which are 0700–0900 LT (the morning peak hours) and 1700–1900 LT (the evening peak hours). First, the predicted commuting flows are obtained for each station and OD using the Prophet model (Taylor and Letham 2018). Second, the changes in flows due to rainfall are estimated based on the observed and predicted values. Third, metro stations are classified according to their usual hourly inbound and outbound flows, which are then validated against their surrounding land-use structure and distribution of point-of-interest (POI). Fourth, the spatial distribution of hourly rainfall is calculated using kriging interpolation. Fifth, the rainfall distribution map, metro network map and commuting flow data are spatially correlated to explore the variations in commuting flows at station and OD levels. (Fig. 3).
The research framework.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
2) Identifying metro station types and estimating OD distances
(i) Classifying metro stations with k-means clustering algorithm
The demand and behavior of commuters during rainfall events vary across areas with different land-use structure (Ma et al. 2018; Li et al. 2020). The resulting changes in inbound commuting flows will also differ at stations within these areas. Therefore, the unsupervised k-means clustering is adopted to classify stations based on their hourly inbound and outbound flows from 0600 to 2200 LT. K-means clustering intends to partition n objects into k clusters in which each object belongs to the cluster with the nearest mean (Hartigan and Wong 1979).
Figure 4 shows the classification results of the metro stations. Figure 4a represents a station in residential areas, characterized by a distinct morning inbound peak and an evening outbound peak. Figure 4b indicates a station in employment areas with a morning outbound peak and an evening inbound peak. Figure 4c exhibits a station in mixed residential and employment areas with inbound and outbound peaks in both morning and evening. Figure 4d shows a station with irregular inbound and outbound flows, mainly located in the airports, tourist attractions, hospitals, and so on. The stations in Fig. 4d are excluded from this study because they are very few in number and only accommodate a small percentage of total commuting flows.
Passenger flow patterns at different types of stations: (a) residential, (b) employment, (c) mixed, and (d) other.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
Based on the surrounding POI distribution obtained from the Baidu Maps, employment stations are further subdivided into commercial and industrial stations. POIs representing commercial areas include shopping malls, supermarkets, banks, office buildings, and so on. POIs representing industrial areas include industrial parks, factories, and so on. Considering the difference in POI size, the weight is set to 1 for commercial stations and 4 for industrial stations according to Hu and Han (2019). In other words, one industrial POI is comparable to four commercial POIs.
Kernel density of (a) commercial and (b) industrial POIs in Shanghai.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
The distribution of the residential stations, commercial stations, mixed stations, and industrial stations is summarized based on the three ring roads in Shanghai (Fig. 6). It is found that 47% of the residential stations are outside the inner ring, 72% of the mixed stations are in the central city near the inner ring, 94% of the industrial stations are between the inner and outer rings, and 86% of the commercial stations are within the inner ring.
Spatial distribution of the four types of stations in Shanghai: (a) point distribution and (b) zonal distribution.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
(ii) Estimating travel distances of OD flows
Travel distance is an important factor influencing the intention and behavior of commuters (Chen et al. 2019; Zhao and Cao 2020). Therefore, the travel distances of commuting OD flows are calculated in batch via the Baidu Maps application programming interface. This will be used to explore the impact of rainfall on commuting flows at different distance levels.
3) Quantifying commuting flow fluctuations during rainfall events
(i) Prediction of baseline commuting flows with no rainfall
(ii) Quantifying the percentage change of commuting flows induced by rainfall
3. Results
a. Predicted baseline commuting flows with no rainfall
Taking 5 June 2020 as an example, Fig. 7 illustrates the predicted baseline flows with no rainfall for the entire metro network, a typical station, and a typical OD. In the morning peak hours (0700–0900 LT), the predicted baseline value of inbound commuting flows is 928 500 as compared with the observed value of 935 400 during rainfall, with a MAPE of 2.23%. In the evening peak hours (1700–1900 LT), the predicted and observed values are 756 900 and 758 300, respectively, with a MAPE of 2.12%. Commuting flows in the morning and evening peak hours increase slightly by 0.75% and 0.18%, respectively, because of rainfall. Meanwhile, well-predicted results are obtained for individual stations and ODs using the Prophet model, too.
Predicted baseline metro flows and observed metro flows on 5 Jun 2020: (a) inbound passenger flows for the entire metro network, (b) inbound passenger flows at the Sanlin Station, and (c) passenger OD flows from the Guilin Road Station to the Sijing Station.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
b. Changes in inbound commuting flows at the four types of stations due to rainfall
1) Overall changes in inbound commuting flows with different rainfall intensity
The percentage change of inbound commuting flows due to rainfall for the four types of stations is calculated during the morning and evening peak hours (Fig. 8). In general, inbound flows tend to decrease with rainfall intensity and vary across station types. During the morning peak hours, inbound flows decrease at most stations except for commercial stations. In particular, residential stations see an average drop of 2.5% during moderate to heavy rainfall and 5.9% during extremely heavy rainfall, the largest reduction in all inbound flows during the morning peak hours. In the evening peak hours, inbound flows at commercial stations are the least variable, with an average drop of 1.4% during moderate to heavy rainfall and 3.4% during extremely heavy rainfall. This is because most commercial stations are located in central city where there are multiple transport connections and high accessibility. As a result, commuters are less affected by rainfall on their way to these stations.
Overall changes in inbound commuting flows with different rainfall intensity (a) in the morning peak hours and (b) in the evening peak hours. Note that the outer shape shows the kernel density estimation of the data distribution, the white dot at the center represents the median, the thick black bar indicates the interquartile range, and the thin black line means the 95% confidence interval.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
2) Stacking effect of inbound commuting flows at the four types of stations
Relevant studies with questionnaires have indicated that rainfall would advance or delay departure time of commuters (Hranac et al. 2006; Cools and Creemers 2013). Figure 9 shows the hourly rainfall and percentage change of inbound commuting flows due to rainfall on 5 June, 6 July, and 15 July 2020. It is found that rainfall generally delays the departure time of commuters, resulting in significant fluctuations in inbound flows that fall first and then rise, that is, a stacking effect. During the three rainfall events, inbound flows drop significantly at 0700 LT and then rise rapidly above normal levels at 0800 and 0900 LT. The same trend is observed between 1700 and 1900 LT.
The stacking effect of inbound commuting flows for (a) 5 Jun 2020 (b) 6 Jul 2020, and (c) 15 Jul 2020.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
The stacking effect of inbound commuting flows varies with station types. In the morning peak hours, the stacking effect is more notable at residential and mixed stations but not at industrial stations. In the evening peak hours, the effect is most significant at commercial stations, followed by industrial and mixed stations.
3) Sensitivity of inbound commuting flows to rainfall at different hours
Based on the data from individual stations in the three rainfall days, regression analysis is performed on the percentage change of inbound commuting flows due to rainfall (dependent variable y) and hourly rainfall intensity (independent variable x) at 0700, 0800, 0900, 1700, 1800, and 1900 LT, respectively, for each of the four station types (Table 2). Some statistically significant relationships are found between the two variables, with residential stations being significant in the morning hours and commercial stations in the evening hours.
Relationships between percentage change of inbound commuting flows and rainfall at different hours. Note that y is the percentage change of inbound commuting flows due to rainfall and x is the hourly rainfall intensity across stations. Statistically significant relationships are shown with the fitted model and R2, and an em dash indicates that the relationship is not statistically significant at the significance level of 0.05.
Figure 10 further displays the percentage change of inbound flows due to rainfall versus rainfall intensity at residential stations in the three morning hours. The inbound flows are more sensitive to rainfall at 0700 LT and less sensitive at 0800 and 0900 LT. This is mainly due to the more flexible departure time for commuters at 0700 LT. Figure 11 shows the sensitivity of inbound flows to rainfall at commercial stations in the three evening hours. The sensitivity is higher at 1700 LT and lower at 1800 and 1900 LT.
Percentage change of inbound flows due to rainfall vs rainfall intensity at residential stations in the three morning hours: (a) 0700, (b) 0800, and (c) 0900 LT.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
Percentage change in inbound flows due to rainfall vs rainfall intensity at commercial stations in the three evening hours: (a) 1700, (b) 1800, and (c) 1900 LT.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
c. Changes in commuting OD flows due to rainfall
1) Changes in commuting OD flows at different distance levels
The travel time distribution of commuting flows in the morning and evening peak hours under normal conditions, that is, without rainfall, is estimated first (Fig. 12a). Commuting flows greater than 100 min are excluded because they only account for 0.02% of the total flows. Following the classification of commuting distance by Huang et al. (2018), it is found that 8.3% of the total commuting flows are within short distance (≤15 min), 65.5% are within medium distance (15–45 min), 18.3% are within long distance (45–60 min), and 7.9% are within extremely long distance (≥60 min).
Changes in commuting OD flows at different distance levels due to rainfall: (a) distribution of OD flows at different travel-time distances under normal conditions, (b) percentage change of OD flows due to rainfall at different travel-time distances, and (c) zonal patterns of variation in commuting OD flows due to rainfall at the four distance levels.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
Meanwhile, the percentage change of commuting flows due to rainfall at different travel-time distances is estimated during the three rainfall events (Fig. 12b). Commuting flows increase by an average of 7.3% within short distance (≤15 min), peaking at nearly 35% in the morning peak hours, and decrease slightly within long distance (>15 min). This is generally consistent with previous questionnaire-based findings (Jain and Singh 2021).
Regional patterns of variation in commuting OD flows due to rainfall at different distance levels are analyzed according to the ring zones in which their departure stations are located (Fig. 12c). It can be seen that the increased short-distance flows mainly occur in the central city within the outer ring. During rainy days, short-distance commuters in the central city who normally take mopeds, bicycles and walking will turn to the metro, which is more punctual and comfortable during the rain (Petrović et al. 2020). This leads to a remarkable increase in metro short-distance flows. However, long-distance flows are not significantly affected by rainfall, as they mostly start or end in the suburbs outside the outer ring, with fewer commuting choices and a higher reliance on the metro.
2) Changes in commuting OD flows due to rainfall across different types of stations
Changes in OD commuting flows due to rainfall across different types of stations are also explored. Four types of commuting flows are selected according to their origins/destinations, including residential → industrial stations and residential → commercial (service) stations in the morning peak hours, industrial → residential stations and commercial (service) → residential stations in the evening peak hours. During the three rainfall events, the overall percentage change of commuting flows for the four types is shown in Fig. 13. Commuting flows for both residential → industrial and residential → commercial (service) stations decrease significantly at 0700 LT and increase at 0800 and 0900 LT. In particular, commuting flows to the industrial areas rise drastically at 0800 and 0900 LT due to their rigorous punctuality regulations. In the evening peak hours, returning commuting flows from the industrial and commercial (service) areas also show a trend of decreasing followed by increasing, with the industrial stations being more notable.
Changes in commuting OD flows across different types of stations during the (a) morning and (b) evening peak hours.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
3) Changes in commuting OD flows on different metro lines
Figure 14 presents the average changes in commuting flows due to rainfall on different metro lines during the three rainfall events. In the morning peak hours, commuting flows decrease significantly at 0700 LT and increase at 0800 and 0900 LT. Lines with significant flow fluctuations include L1, L7, L8, and L9. The departure stations of these lines are mostly located in large residential areas, which are heavily dependent on the metro. Rainfall can cause surges and overloads of commuting flows on these lines at certain hours. In the evening peak hours, metro commuting flows tend to decrease at 1700 LT and increase at 1800 and 1900 LT. The lines with notable flow fluctuations are mainly those with departure stations in large industrial parks and commercial centers. Relative to the morning peak hours, commuting flows in the evening peak hours fluctuate less.
Changes in commuting OD flows on metro limes during the morning and evening peak hours.
Citation: Weather, Climate, and Society 14, 3; 10.1175/WCAS-D-21-0167.1
4. Discussions
a. Policy implications
As compared with previous studies that analyzed the relationship between rainfall and commuting at daily scale, this paper examines the commuting flows at different times in the morning and evening peak hours, as well as their spatial distribution from metro stations during the rainfall events. This helps to investigate the spatiotemporal patterns of commuting flows in more detail. From these flow patterns, spatiotemporal differences in the demand and behavior of commuters can be further explored. It is generally accepted that the total daily commuting flows are regular and less affected by rainfall, which is also verified in section 3a. However, at finer scales, significant surges and stacking effects of commuting flows can be observed at certain hours and stations.
Based on rainfall forecast information with intensity, timing and extent, the method established in this study could be further used to estimate the potential impact of rainfall on commuting flows in advance, identify the stations and lines that are most sensitive to rainfall at different times, and make specific preparations and responses. This is of importance for local agencies to make reasonable response plans for various rainfall scenarios.
For those stations and lines that are sensitive to rainfall, mitigation measures, including regional multiple transport modes integration and rational land-use planning, are necessary to reduce the negative impact on local commuting. For example, in some areas, commuters are used to ride mopeds and shared bicycles to metro stations, which are susceptible to rainfall. Rainfall can result in their delays in departures from home and consequent stacking of inbound flows at metro stations. Therefore, additional feeder modes, such as shuttles to metro stations, can be temporarily deployed during peak hours on rainfall days to reduce the impact on commuting flows. In addition, some large residential areas located in the near suburbs have fewer commuting options and are more reliant on the metro. The metro stations and lines in these areas are already close to capacity during commuting hours. Rainfall can exacerbate the flow fluctuations and overloads at certain times. It is therefore necessary to relieve the pressure on metro by enriching other commuting modes, such as surface public transport. In the long term, commuting needs and pressures need to be alleviated at source by optimizing the local land-use structure, increasing employment opportunities, and reducing the job-housing imbalance.
b. Uncertainties
There are some uncertainties in the results. First, the time, duration, intensity, and spatial extent of rainfall events vary considerably. Our results are only based on three rainfall events, which might cause certain uncertainties. Adequate rainfall event samples are needed to obtain more reliable results. In addition, detailed rainfall types can be distinguished from these samples and used to analyze their respective impact on commuting flows. Second, the built environment around metro stations is highly spatially heterogeneous, in which various factors may affect commuting choices (Thorhauge et al. 2016; Markolf et al. 2019). We roughly classify the stations according to their inbound/outbound flow patterns and surrounding land-use structure, without considering the various built-up environment factors. Third, feeder-related factors between home/work locations and metro stations also influence the decisions and actions of commuters (Griffin and Sener 2016; Fan and Zheng 2020), such as distance levels, available feeder modes, and available walkways. These feeder-related factors should be incorporated in future studies to analyze their impact on commuting flows. Fourth, the ground traffic congestion resulting from rainfall may also lead to delayed departure time of commuters. In the future, new methods that integrate predictions of ground congestion and metro congestion during rainfall can provide timely decision support for urban traffic congestion management (Yan et al. 2020).
5. Conclusions
Taking Shanghai as an example, the Prophet time series model is used to derive the baseline metro flows with no rainfall, which are then compared with the observed metro flows during rainfall to quantify the spatiotemporal variations of metro commuting flows due to rainfall. The main results are as follows:
-
In general, inbound commuting flows at metro stations tend to decrease with hourly rainfall, varying across station types. In the morning peak hours, residential stations see an average drop of 2.5% during moderate to heavy rainfall and 5.9% during extremely heavy rainfall, the largest reduction in inbound flows. In the evening peak hours, inbound flows at commercial stations exhibits the least variation, with an average drop of 1.4% during moderate to heavy rainfall and 3.4% during extremely heavy rainfall. The departure time of commuters is usually delayed due to rainfall, resulting in a significant stacking effect of inbound flows at metro stations, showing a pattern of falling first followed by a rise. The stacking effect is notable at residential stations in the morning peak hours and at commercial and industrial stations in the evening peak hours. The sensitivity of commuting flows to rainfall varies at different times—high at 0700 and 1700 LT and low at 0800, 0900, 1800, and 1900 LT—because of the different levels of flexibility of departure time.
-
Rainfall has a greater impact on short distance commuting flows. Commuting OD flows within short distance (≤15 min) increase by an average of 7.3%, peaking at nearly 35%, while the flows within long distance (>15 min) decrease slightly. Commuting OD flows between residential and industrial areas are more affected by rainfall than those between residential and commercial (service) areas, showing a greater fluctuation of falling followed by a rise. The sensitivity of commuting OD flows to rainfall varies across metro lines. The departure stations of rainfall-sensitive lines are mostly distributed in large residential areas in the morning peak hours, and in large industrial parks and commercial centers in the evening peak hours. Commuting flows in the evening peak hours are less fluctuating than in the morning peak hours.
-
This paper explores the spatiotemporal variations of commuting flows due to rainfall at a finer scale, which provides a sound basis for spatial and temporal response strategies. Our findings suggest that there are significant surges and stacking effects of commuting flows at certain times and areas induced by rainfall. Attention should be paid to these fluctuations and overloads for spatially and temporally aware decision-making.
Acknowledgments.
This work was supported by the National Natural Science Foundation of China (Grants 41771540 and 42171080) and the National Key Research and Development Program of China (Grants 2017YFC1503001 and 2019YFB2101600).
Data availability statement.
Hourly rainfall data used during this study are openly crawled from the official website (http://bmxx.swj.sh.gov.cn/shfx) established by the Shanghai Water Authority (Shanghai Municipal Oceanic Bureau). Because of its proprietary nature, metro flow data used during this study cannot be made openly available. Further information about the data and conditions for access is available from the authors.
REFERENCES
Arana, P., S. Cabezudo, and M. Peñalba, 2014: Influence of weather conditions on transit ridership: A statistical study using data from Smartcards. Transp. Res., 59A, 1–12, https://doi.org/10.1016/j.tra.2013.10.019.
Böcker, L., M. Dijst, and J. Prillwitz, 2013: Impact of everyday weather on individual daily travel behaviours in perspective: A literature review. Transp. Rev., 33, 71–91, https://doi.org/10.1080/01441647.2012.747114.
Chen, E., Z. Ye, and H. Bi, 2019: Incorporating smart card data in spatio-temporal analysis of metro travel distances. Sustainability, 11, 7069, https://doi.org/10.3390/su11247069.
Cools, M., and L. Creemers, 2013: The dual role of weather forecasts on changes in activity-travel behavior. J. Transp. Geogr., 28, 167–175, https://doi.org/10.1016/j.jtrangeo.2012.11.002.
Diab, E., and A. Shalaby, 2020: Metro transit system resilience: Understanding the impacts of outdoor tracks and weather conditions on metro system interruptions. Int. J. Sustainable Transp., 14, 657–670, https://doi.org/10.1080/15568318.2019.1600174.
Fan, Y., and S. Zheng, 2020: Dockless bike sharing alleviates road congestion by complementing subway travel: Evidence from Beijing. Cities, 107, 102895, https://doi.org/10.1016/j.cities.2020.102895.
Griffin, G. P., and I. N. Sener, 2016: Planning for bike share connectivity to rail transit. J. Public Transp., 19, 1–22, https://doi.org/10.5038/2375-0901.19.2.1.
Hartigan, J. A., and M. A. Wong, 1979: Algorithm AS 136: A k-means clustering algorithm. J. Roy. Stat. Soc. Ser., 28C, 100–108, http://doi.org/10.2307/2346830.
Hranac, R., E. D. Sterzin, D. Krechmer, H. Rakha, and M. Farzaneh, 2006: Empirical studies on traffic flow in inclement weather. FHWA Tech. Rep. FHWA-HOP-07-073, 114 pp.
Hu, Y., and Y. Han, 2019: Identification of urban functional areas based on POI data: A case study of the Guangzhou economic and technological development zone. Sustainability, 11, 1385, https://doi.org/10.3390/su11051385.
Huang, J., D. Levinson, J. Wang, J. Zhou, and Z. Wang, 2018: Tracking job and housing dynamics with smartcard data. Proc. Natl. Acad. Sci. USA, 115, 12 710–12 715, https://doi.org/10.1073/pnas.1815928115.
IPCC, 2021: Climate Change 2021: The Physical Science Basis. V. Masson-Delmotte et al., Eds., Cambridge University Press, 3949 pp.
Jain, D., and S. Singh, 2021: Adaptation of trips by metro rail users at two stations in extreme weather conditions: Delhi. Urban Climate, 36, 100766, https://doi.org/10.1016/j.uclim.2020.100766.
Kalkstein, A. J., M. Kuby, D. Gerrity, and J. J. Clancy, 2009: An analysis of air mass effects on rail ridership in three US cities. J. Transp. Geogr., 17, 198–207, https://doi.org/10.1016/j.jtrangeo.2008.07.003.
Kashfi, S. A., J. M. Bunker, and T. Yigitcanlar, 2016: Modelling and analysing effects of complex seasonality and weather on an area’s daily transit ridership rate. J. Transp. Geogr., 54, 310–324, https://doi.org/10.1016/j.jtrangeo.2016.06.018.
Li, J., X. Li, D. Chen, and L. Godding, 2018: Assessment of metro ridership fluctuation caused by weather conditions in Asian context: Using archived weather and ridership data in Nanjing. J. Transp. Geogr., 66, 356–368, https://doi.org/10.1016/j.jtrangeo.2017.10.023.
Li, S., D. Lyu, G. Huang, X. Zhang, F. Gao, Y. Chen, and X. Liu, 2020: Spatially varying impacts of built environment factors on rail transit ridership at station level: A case study in Guangzhou, China. J. Transp. Geogr., 82, 102631, https://doi.org/10.1016/j.jtrangeo.2019.102631.
Liang, P., and Y. Ding, 2017: The long-term variation of extreme heavy precipitation and its link to urbanization effects in Shanghai during 1916–2014. Adv. Atmos. Sci., 34, 321–334, https://doi.org/10.1007/s00376-016-6120-0.
Ma, X., J. Zhang, C. Ding, and Y. Wang, 2018: A geographically and temporally weighted regression model to explore the spatiotemporal influence of built environment on transit ridership. Comput. Environ. Urban Syst., 70, 113–124, https://doi.org/10.1016/j.compenvurbsys.2018.03.001.
Markolf, S. A., C. Hoehne, A. Fraser, M. V. Chester, and B. S. Underwood, 2019: Transportation resilience to climate change and extreme weather events—Beyond risk and robustness. Transp. Policy, 74, 174–186, https://doi.org/10.1016/j.tranpol.2018.11.003.
Najafabadi, S., A. Hamidi, M. Allahviranloo, and N. Devineni, 2019: Does demand for subway ridership in Manhattan depend on the rainfall events? Transp. Policy, 74, 201–213, https://doi.org/10.1016/j.tranpol.2018.11.019.
Pendergrass, A. G., and D. L. Hartmann, 2014: Changes in the distribution of rain frequency and intensity in response to global warming. J. Climate, 27, 8372–8383, https://doi.org/10.1175/JCLI-D-14-00183.1.
Petrović, D., I. Ivanović, V. Đorić, and J. Jović, 2020: Impact of weather conditions on travel demand—The most common research methods and applied models. Promet Traffic Transp., 32, 711–725, https://doi.org/10.7307/ptt.v32i5.3499.
Sillmann, J., V. V. Kharin, F. Zwiers, X. Zhang, and D. Bronaugh, 2013: Climate extremes indices in the CMIP5 multimodel ensemble: Part 2. Future climate projections. J. Geophys. Res., 118, 2473–2493, https://doi.org/10.1002/jgrd.50188.
Su, S., T. Tang, and Y. Wang, 2016: Evaluation of strategies to reducing traction energy consumption of metro systems using an optimal train control simulation model. Energies, 9, 105, https://doi.org/10.3390/en9020105.
Tang, H., Y. Zhang, H. Liu, D. Kong, and Y. Ge, 2018: Impact analysis of rainfall on Beijing subway transit ridership with heterogeneous characteristics. Proc. 17th COTA Int. Conf. of Transportation Professionals, Shanghai, China, Chinese Overseas Transportation Association, 236–247.
Tao, S., J. Corcoran, M. Hickman, and R. Stimson, 2016: The influence of weather on local geographical patterns of bus usage. J. Transp. Geogr., 54, 66–80, https://doi.org/10.1016/j.jtrangeo.2016.05.009.
Tao, S., J. Corcoran, F. Rowe, and M. Hickman, 2018: To travel or not to travel: ‘Weather’ is the question. Modelling the effect of local weather conditions on bus ridership. Transp. Res., 86C, 147–167, https://doi.org/10.1016/j.trc.2017.11.005.
Taylor, S. J., and B. Letham, 2018: Forecasting at scale. Amer. Stat., 72, 37–45, https://doi.org/10.1080/00031305.2017.1380080.
Thorhauge, M., E. Cherchi, and J. Rich, 2016: How flexible is flexible? Accounting for the effect of rescheduling possibilities in choice of departure time for work trips. Transp. Res., 86A, 177–193, https://doi.org/10.1016/j.tra.2016.02.006.
Wang, M., 2018: Shanghai Flood Control Manual. Fudan University Press, 604 pp.
Wei, M., Y. Liu, T. Sigler, and J. Corcoran, 2021: Unpacking the weather-transit ridership relationship using big data in Brisbane and beyond. Big Data Applications in Geography and Planning, M. Birkin et al., Eds., Edward Elgar, 245–255.
Wu, J., and H. Liao, 2020: Weather, travel mode choice, and impacts on subway ridership in Beijing. Transp. Res., 135A, 264–279, https://doi.org/10.1016/j.tra.2020.03.020.
Yan, C., X. Wei, X. Liu, Z. Liu, J. Guo, Z. Li, Y. Lu, and X. He, 2020: A new method for real-time evaluation of urban traffic congestion: A case study in Xi’an, China. Geocarto Int., 35, 1033–1048, https://doi.org/10.1080/10106049.2018.1552325.
Yang, S., G. Yin, X. Shi, H. Liu, and Y. Zou, 2016: Modeling the adverse impact of rainstorms on a regional transport network. Int. J. Disaster Risk Sci., 7, 77–87, https://doi.org/10.1007/s13753-016-0082-9.
Zanni, A. M., and T. J. Ryley, 2015: The impact of extreme weather conditions on long distance travel behaviour. Transp. Res., 77A, 305–319, https://doi.org/10.1016/j.tra.2015.04.025.
Zhao, P., and Y. Cao, 2020: Commuting inequity and its determinants in Shanghai: New findings from big-data analytics. Transp. Policy, 92, 20–37, https://doi.org/10.1016/j.tranpol.2020.03.006.
Zhou, M., D. Wang, Q. Li, Y. Yue, W. Tu, and R. Cao, 2017: Impacts of weather on public transport ridership: Results from mining data from different sources. Transp. Res., 75C, 17–29, https://doi.org/10.1016/j.trc.2016.12.001.
Zhou, Y., Z. Li, Y. Meng, Z. Li, and M. Zhong, 2021: Analyzing spatio-temporal impacts of extreme rainfall events on metro ridership characteristics. Physica A, 577, 126053, https://doi.org/10.1016/j.physa.2021.126053.
