1 Introduction

Multimodal public transportation is generally believed to promote job/work accessibility, equity, and commercial prosperity. This study investigates the potential transfer of dockless shared bikes (DSB) to urban rail transit (URT) in Shenzhen, China. In China, especially in large cities, the URT system is considered better for longer-distance daily commuting to work/school because of its punctuality and stability. The Shenzhen 2019 Household Travel Survey [1] shows that the average travel distance for subway trips is 14.7 km and the average subway travel time is 54 minutes. Since Shenzhen opened its first metro line in 2004, with the rapid city development resulting from the Open and Reform Policy, Shenzhen metro miles have increased rapidly. In 2020 alone, Shenzhen opened six new metro lines, which is an increase of 107 km of subway lines for Shenzhen [2]. At the end of 2021, the average daily passenger transport intensity of Shenzhen was 14.5 k/km, ranking first in China, and the subway constitutes more than 60% of people’s total travel, meaning the subway plays a very important role in people’s daily travel lives [2].

As one of the most environmentally friendly transit modes, compared with docked shared bikes, DSB provide people with more flexibility for rent and return, making them more accessible and much easier to use. With the rapid development of DSB since 2017, in China’s large cities such as Beijing, Shanghai, Guangzhou, and Shenzhen, data show that this travel mode greatly encourages people to use it as a transfer mode from/to subway lines for the “last mile” commute [3, 4]. In a study by Du and Cheng [5], a survey of 4939 respondents was conducted in a Chinese city, Nanjing, and the results show that 51.13% of the users ride a DSB to subway stations for transfer.

Promoting DSB is very important as it helps to divert travelers to a clean and non-automobile transit option, and predicting its demand is very important to balance the pick-up and drop-off demand to improve the accessibility and availability of DSB, especially at service hot spots such as malls and public transport stations. Imbalanced demand in hot spots for DSB can lead to overcrowding, with the number of drop-off bikes exceeding the local parking capacity, or vacancies in popular pick-up locations with few bikes available. These situations often require bike-delivery services for demand rebalancing, which can increase bike loss rates due to damages incurred during the delivery process. Furthermore, this can decrease the efficiency of the transportation system, as the delivery process may block bike pick-up and drop-off locations and cause traffic disruptions in the surrounding areas. Additionally, the delivery process is typically not environmentally friendly, as it often involves internal combustion engine pick-up trucks or minivans.

To promote the usage of shared bikes for first- and last-mile trips and further promote subway usage, integration between the two trip modes’ networks is important. To achieve this integration, the following methods can be considered:

  1. 1.

    Demand complementarity study: Conducting a demand complementarity study can help identify areas where shared bikes and subways can work together to meet commuters' needs. For example, areas with high subway demand but limited bike infrastructure may benefit from targeted bike-sharing programs.

  2. 2.

    Station and bike hub integration: Integrating bike-sharing hubs with subway stations can encourage commuters to switch between modes of transportation. Placing bike-sharing stations in close proximity to subway stations can make it easier for commuters to use bikes for the first and last mile of their trip.

  3. 3.

    Joint route planning: Incorporating bike-sharing stations into subway route planning can help commuters find the most efficient and convenient way to reach their destination. For example, a commuter may choose to take the subway to a station close to their destination and then use a shared bike to complete the last mile of their trip.

  4. 4.

    Real-time information: Providing real-time information on bike and subway availability can help commuters plan their trips more effectively. This can include information on bike availability, subway delays, and real-time route planning.

  5. 5.

    Incentives: Offering incentives to commuters who use both shared bikes and subways can help encourage more people to try this combined mode of transportation. For example, offering discounted fares or free rides for bike-sharing users who also use the subway can encourage more people to use both modes of transportation.

Thus, it is important to understand the DSB demand and predict its demand in advance for all the methods mentioned above. DSB often serves as a first-/last-mile transfer for various trip purposes such as commuting and recreation, especially around subway stations. However, the DSB usage (i.e., the pick-up and drop-off demand for DSB) varies across time of day (e.g., morning peak vs. mid-day off-peak vs. evening peak) and day of the week (e.g., workdays vs. weekends). This variability is nontrivial, and sometimes can be huge across locations (city hot spots). Understanding the pick-up and drop-off demand and associated temporal–spatial variability facilitates the optimization of the bike delivery process across time and space, thus reducing the bike rebalancing service needs and supporting the local bike management process in terms of land use and labor needs.

The prediction of both pick-up and drop-off demand needs to be performed for imbalanced demand gap identification and thus for better DSB management. URT stations are often hot spots for DSB pick-up and drop-off, and it is important to understand the DSB usage at URT stations for promoting DSB ridership and multimodal transit. Predicting the pick-up and drop-off demand at URT stations helps mitigate local shared bike parking congestion problems and increases DSB accessibility.

In the era of big data and machine learning, traditional traffic forecasting tasks such as traffic flow prediction have already seen substantial growth in the usage of deep learning models, because these models demonstrate better performance on big data. Thus, it is both important and necessary to use deep learning methods for the prediction of DSB demand. In this study, the most recent deep learning methods used in other prediction-related tasks (mostly traffic flow prediction) [6] are adopted and implemented for DSB demand prediction, which will hopefully serve as the first step in tacking the problems both the government and the operating companies are facing for the management of DSB.

2 Literature Review

To promote first- and last-mile trip bike usage, thus further promoting subway usage, the first step is to predict bike usage around subway stations. The following section provides a literature review on demand prediction methods.

For traditional travel modes (cars, transit, traditional bikes, etc.), when forecasting this demand, there is a trend for researchers to move away from conventional travel demand models, specifically the four-step model and the activity-based model. These traditional models necessitate substantial effort in collecting comprehensive data, including travel survey data and land use data. However, recent advancements in machine learning and the proliferation of big data have led scholars to adopt data-driven approaches for traffic demand prediction. In particular with the emergence of car-sharing and bike-sharing systems, traditional survey data-supported models become infeasible. Thus, the evolution of methods employed for traffic demand prediction has transitioned away from early models grounded in economic theory, such as discrete choice models, which offer theoretical explainability when thoughtfully designed with well-crafted features. Subsequently, researchers have incorporated machine learning techniques for demand prediction, including regression-based models, tree-based models, and support vector machines. Most recently, the field has seen the integration of cutting-edge deep learning methods into the repertoire of predictive tools. As a correlated time series prediction task, long short-term memory (LSTM) or gated recurrent unit (GRU) approaches can be used to model the temporal dynamics in the travel demand data [7,8,9,10]. Some efforts employ temporal convolutional networks (TCN), which enable the modeling of very long sequences with less time [11,12,13]. To capture the spatial interactions among different traffic series, these efforts [14,15,16] formulate the traffic forecasting problem on a graph and utilize the spectral graph convolutional network (GCN). Then the diffusion convolutional recurrent neural network (DCRNN) [17] reformulates the spatial dependence of traffic as a diffusion process and extends the previous GCN [18,19,20] to a directed graph. Graph WaveNet [21] combines GCN with dilated causal convolution networks for saving computation cost and uses a self-adaptive adjacency matrix as a complement for the predefined adjacent matrix to capture spatial correlations. More recent works such as ASTGCN [22], STSGCN [23] and GMAN [24] attention-based spatial-temporal graph convolutional network (ASTGCN) [17], spatial-temporal synchronous graph convolutional network (STSGCN) [18], and graph multi-attention network (GMAN) [19] add further complicated spatial and temporal attention mechanisms with GCN to capture the dynamic spatial and temporal correlations. However, these methods can only capture shared patterns among all traffic series and still rely on the predefined spatial connection graph.

Many studies have already been conducted for DSB demand analysis. DSB demand exhibits distinct temporal and spatial patterns. On one hand, weekdays and weekends, and peak and non-peak hours, show different patterns, with substantial spatial variability. On the other hand, a considerable portion of shared bikes are parked in public urban centers, particularly around transit and subway stations, indicating distinct spatial patterns [17, 25,26,27,28]. Factors affecting shared bike riding can be roughly categorized as demand factors, supply factors, and other environmental factors. Demand factors such as population [29, 30], supply factors such as traffic accessibility, infrastructure conditions, parking conditions, and number of points of interest (POI) [20, 29, 31–33], and other exogenous environmental factors such as the weather and temperature [19, 34,35,36,37] are all shown to affect the daily volume of regional bike sharing. For example, in a study by Kim [35], it was found that when the temperature was above 30°C, shared bike usage decreased.

For shared bike demand predictions, deep learning methods have been used in recent years. Deep learning methods are most widely used for traffic flow prediction tasks, and there are already review papers that summarize the deep learning methods used in traffic flow prediction [6]. An LSTM model was used in a 2018 DSB study in the Nanjing downtown area [18], and the results show that LSTM performs better than the conventional machine learning models such as autoregressive integrated moving average (ARIMA), extreme gradient boosting (XGBoost), and support vector machine (SVM). However, the data used in the study only include one of the operating companies for the study area, which might introduce some bias for the whole city’s demand analysis.

Both Ai et al. [38] and Wang et al. [39] used a conv-LSTM model to forecast the short-term distribution of shared bike programs in Chengdu, and demonstrated the applicability of the conv-LSTM model for shared bike demand prediction in terms of promoting prediction accuracy. A TCN model was also used to predict the DSB demand by clustered traffic analysis zones (TAZs) in a 2020 study [40], in which the results also indicated that the TCN performed better than the conventional statistical models. In other works [41, 42], graph structures of travel flow spatial and temporal patterns are used for interpreting urban dynamics as well as for traffic demand prediction. However, there is no study with extensive experiments comparing the most recent deep learning methods such as DCRNN, Graph WaveNet, and adaptive graph convolutional recurrent network (AGCRN).

As mentioned above, although DSB are a very important means of connection to complete people’s first and last miles for subway rides, there are not many studies which focus on the evolution of DSB demand around subway stations with large datasets using the most up-to-date deep learning methods. This study will first dive into the temporal and spatial analysis of DSB usage in Shenzhen with city-wide data, in particular the potential transfer bikes around subway stations, and then use the most recent deep learning methods, which include LSTM, DCRNN, Graph WaveNet, and AGCRN, to predict DBS demand around subway stations, thus potentially providing operating companies and traffic management agencies with deeper insight into DSB supply and demand balance regulations.

3 Methodology

This methodology section will first introduce the processing details of all the data used in this study (metro usage and bike usage data all included), then the bike usage around subway analysis is introduced, and finally, the demand prediction methods used are shown together with the forecasting results.

In order to determine the potential transfer of bikes and the subway, we obtained the subway and bike usage data from two different sources. Firstly, the subway data are processed and described in the following.

3.1 Subway Data Analysis

The subway data are from August 31, 2018, to September 1, 2018, with a total of 1,337,000 records. The data are preprocessed to obtain the subway station volume. By a day granularity, Fig. 1 shows the subway stations plotted in QGIS with Amap tiles shown as the bottom layer and the circle radius representing the volume density. Ranked as #1, Shenzhen North Station is a crucial transportation hub that serves as an interchange station for MTR (Shenzhen) Line 4, Shenzhen Metro Line 5, and Line 6. While MTR (Shenzhen) Line 4 and Shenzhen Metro Line 6 stations are elevated, the station for Shenzhen Metro Line 5 is underground. The station connects passengers to the Shenzhen North Railway Station and plays a significant role in safely transporting daily commuters during peak hours, making it the station with the largest passenger volume. Figure 2 shows the top 10 highest-volume subway stations plotted in QGIS.

Fig. 1
figure 1

Subway stations with volume shown with GIS

Full size image
Fig. 2
figure 2

Top 10 highest-volume subway stations shown in GIS

Full size image

3.2 Bike Data Analysis

This study examines the DSB market in Shenzhen, China, which is dominated by three primary commercial operators: Meituan, Hello, and Didi bikes. Meituan has a fleet of 262,000 operating bikes, while Hello and Didi bikes operate 100,000 and 59,000 bikes, respectively, resulting in a total of 420,000 DSB operating in the city. Data collected from April 1, 2021, to May 28, 2021, were obtained from all three companies via the Shenzhen open public data platform application programming interface (API). The data set comprises starting and ending location coordinates and corresponding time stamps for each trip taken.

To facilitate mapping and trip distance calculation using OpenStreetMap, the trip start and end location coordinates were converted from the original gcj02 projection system to the wgs08 system. The process of transforming the coordinates is demonstrated in Fig. 3, which displays a sample of data points before and after the transformation.

Fig. 3
figure 3

Example results of coordinate transformation

Full size image

This study implemented two criteria for data selection: boundary selection and trip time/distance calculation. First, trips with start or end locations outside of the Shenzhen boundary limits were excluded. Second, trip duration was calculated as the difference between trip start and end time, and any trips lasting more than 1 hour were excluded from the analysis. The trip distance was calculated using the haversine formula based on start and end location coordinates, with any trips exceeding 5 km also removed from the data set. Figure 4 displays the data retention rates after processing for each of the 58 days included in this study. It indicates that April 3–7 had comparatively lower retention rates than the other days in April, likely due to the national holiday observed on April 3, 4, and 5. As some individuals may have extended their holiday into the following week, longer trip durations and distances were likely more prevalent, resulting in lower retention rates for these days. Figure 4 also shows lower retention rates for May 1–3, because those days are also national holidays, and April 25 (Sunday) and May 8 (Saturday) are adjusted to be usual working days.

Fig. 4
figure 4

April and May DSB data retention rates

Full size image

The Shenzhen metro system is still rapidly expanding, with many lines currently in the planning or construction phase, and only data from 2021 DSB usage were available for analysis in this paper. Therefore, the left side of Fig. 5 depicts a map of Shenzhen showing the subway stations available as of 2021. In total, there were 11 subway lines with 236 stations (with transfer stations counted as one station) in operation during this period. To determine the number of potentially connected trips with subway rides, we employed the catchment area method. In this study, station locations were defined as the latitude and longitude coordinates of station centers (depicted as red points on the right side of Fig. 4). The distance between station exits and centers ranged from 30 to 130 m for most stations. For example, Fig. 5 shows two sample stations in Shenzhen, Xili and Renmin South, and their respective distances from station exits F and B. Based on our analysis, we selected a station radius of 150 m as the catchment area for this study.

Fig. 5
figure 5

May 2021 Shenzhen Metro stations with Xili and Renmin south stations as examples

Full size image

As an illustration of trip time and distance distribution, we used a typical weekday (April 14, 2021) and plotted all-day trips, pick-up trips, and drop-off trips in Fig. 5, after applying the data filtering process. Out of a total of 2,562,188 DSB trips taken on April 14, 2021, we selected 2,488,516 trips within a 150-m radius catchment area of subway stations. Specifically, we identified 104,054 pick-up trips and 102,609 drop-off trips within this catchment area. The time and distance distribution for both pick-up and drop-off trips at subway stations were very similar, with no significant differences observed. The majority of trips were between 0.3 and 2 km in distance and lasted between 5 and 20 minutes (Fig. 6).

Fig. 6
figure 6

April 4–14, 2021 DSB trips distance and time distribution

Full size image

3.3 Use of Shared Bikes for Subway Transfer

Figure 7 illustrates the distribution of bike demand and the volume of Shenzhen subway stations. There is a strong correlation between the high-demand spots for bikes and the subway. The areas with high-density bike usage are mostly located around the URT stations, and the usage of DSB is also strongly related to the surrounding land use, with denser work and residential areas showing higher demand. These findings suggest that metropolitan areas surrounding metro stations, which have a high combination of various land uses and a significant proportion of industrial land use, tend to have a more integrated utilization pattern. The absence of bike-sharing coverage at the terminal points of subway lines 3, 4, and 10 (blue circle area in Fig. 7) can primarily be attributed to their noncentral location and the surrounding developing areas.

Fig. 7
figure 7

Visualization of DSB and subway usage on amap tile map

Full size image

In the top 20 bike volume around stations from 2021 data, seven stations (bold in Table 1) are new stations built after 2018, among which the Science Museum Station is ranked first, which implies the potential effect of a subway on bike usage as the first/last mile.

Table 1 Top 20 ranked total bike volume
Full size table

3.4 DSB Supply Imbalance Around Subway Stations

Figure 8 presents the hourly average start–end imbalance number (calculated as start bike number minus end bike number) for all stations. The majority of stations exhibit a nonzero start-end number, indicating potential bike shortages (when positive) or surpluses (when negative). Notably, the imbalance issues vary significantly among subway stations, with the largest number being 16.8 for station 235 (Houting) and the smallest number being −17.7 for station number 27 (Hubei). Figures 9 and 10 depict the maximum and minimum start-end numbers for all subway stations in April and May. Figure 9 shows that some stations experience a difference of several hundred start trips compared to end trips during the 58-day period, with the highest number being 428 for station 215 (Gangtou), indicating a potential shortage. Figure 10 shows that for certain stations, end trips exceed start trips by several hundred, with station 207 (Wenjin) having the lowest number of −369, indicating a considerable surplus. These findings suggest that the issue of supply imbalance is station-specific and highlight the need for targeted intervention strategies.

Fig. 8
figure 8

Average DSB usage imbalance across the day (per hour)

Full size image
Fig. 9
figure 9

Maximum DSB usage imbalance across the day (per hour)

Full size image
Fig. 10
figure 10

Minimum DSB supply imbalance across the day (per hour)

Full size image

3.5 DSB Demand Prediction

To promote the usage of subways, enhance their integration with DSB, and tackle the imbalance problem, accurate demand prediction is necessary. This section presents the results of DSB demand prediction. The long short-term memory (LSTM) [27], spatiotemporal neural network (STNN) [28], diffusion convolutional recurrent neural network (DCRNN) [17], Graph WaveNet [29], and adaptive graph convolutional recurrent network (AGCRN) [43] are all used in this study. AGCRN is a cutting-edge method for predicting traffic demand in urban areas, which is a deep learning-based approach that leverages the power of graph convolutional networks (GCNs) and recurrent neural networks (RNNs) to model the spatial and temporal dependencies of traffic demand. The AGCRN method is designed to handle the dynamic nature of traffic patterns in urban areas. It is able to adaptively learn the spatial and temporal characteristics of the traffic network and generate accurate predictions for future traffic demand. The graph convolutional network is responsible for modeling the spatial dependencies of traffic flows and is used to aggregate information from the neighboring nodes and generate node embeddings, which are then fed into the RNN. The recurrent neural network is responsible for modeling the temporal dependencies of traffic flows. It uses a sequence of node embeddings generated by the GCN layers as input and predicts the traffic demand for each intersection at the next time step. The attention mechanism is also incorporated to enhance the model's capability to attend to relevant spatial and temporal features. The AGCRN method has been shown to outperform traditional time-series forecasting methods and other deep learning-based methods in predicting traffic demand. Its ability to adaptively learn the spatial and temporal characteristics of the traffic network and generate accurate predictions makes it a valuable tool for transportation planning and management in urban areas. Because the AGCRN method also incorporates an attention mechanism to enable the model to attend to relevant spatial and temporal features, it can help the model to focus on important bike stations and time periods, improving the accuracy of the demand predictions. Its ability to adaptively learn the spatial and temporal patterns of bike usage and make accurate predictions can help bike-sharing companies to optimize their operations, such as bike deployment and redistribution, to meet the demand and improve the overall user experience. In the following, we present the results of the DSB drop-off and pick-up demand predictions separately. We set the historical demand length to 12 hours, and the forecast length is also set to 12 hours, which means we use the previous 12 hours of demand to predict the next 12 hours of demand. In our experiment, we used 70% of the data for training, 10% for validation, and the remaining 20% for testing.

Three metrics frequently used in model evaluations are utilized in our study, i.e., mean absolute error (MAE), root-mean-square error (RMSE), and mean absolute percentage error (MAPE). The following four methodsFootnote 1 are used for DSB demand prediction comparison with AGCRN in this study:

  1. 1.

    Long short-term memory (LSTM) [27] encoder historical sequences with LSTM.

  2. 2.

    Spatiotemporal neural network (STNN) [28] learns dependencies through a structured latent dynamical component and makes use of a decoder mechanism for prediction.

  3. 3.

    Diffusion convolutional recurrent neural network (DCRNN) [17] integrates diffusion convolution with sequence-to-sequence architecture.

  4. 4.

    Graph WaveNet [29] combines adaptive graph convolution with dilated casual convolution.

Table 2 shows the prediction results in terms of the three metrics of the five different methods for both the pick-up and drop-off demand. For pick-up demand prediction, AGCRN performs better than the other four methods in terms of MAE, RMSE, and MAPE. For the drop-off demand prediction task, AGCRN can also be considered best for its lowest MAE and MAPE, and its RMSE is the second smallest (8.5530) compared with the other four methods. For all the methods with graphs, we implemented a Shenzhen metro station adjacency matrix graph, where if two stations are connected, then a weight of 1 is assigned to the vertex, and otherwise, 0 is assigned. From Table 2 we see that AGCRN has the best performance for DBS demand prediction in this study. LSTM and STNN are relatively simpler deep learning methods that only consider the temporal correlations; thus, their performance is not as good as AGCRN. DCRNN based on RNN and diffusion GCN does not achieve the best results in terms of the MAE, RMSE, and MAPE scores, because RNN cannot capture long-term temporal correlations. Graph WaveNet, which uses TCN along the temporal dimension to capture the temporal correlations, obtains relatively good results, but still not the best in this study. AGCRN extends GCN with self-learned adjacency matrices and obtains the best results here, showing that self-learned adjacency matrices are effective for the traffic forecasting task here, potentially because it captures the hidden relationship between subway stations for shared bike usage around these stations.

Table 2 Prediction performance comparison of five methods
Full size table

4 Demand Prediction Results Discussion

This section will show the demand prediction results using AGCRN in terms of two aspects: the temporal aspect and the spatial aspect.

Temporal results are shown for fixed stations. Ten stations with high DSB demand (Table 3) are selected. The pick-up and drop-off demand prediction results across time for these 10 stations are shown in Figs. 7 and 8. For most of the station predictions, the method captures the trends well, such as the AM and PM peaks for both the pick-up and the drop-off demand. However, with regard to the demand peaks for some stations (e.g., pick-up demand peaks for station number 173, drop-off demand peaks for station numbers 38, 22, 214, 21), although AGCRN outperforms the other four methods, it still does not capture the peak demand well, mostly underestimating the high demand and showing an obvious gap for the highest demand across time for most stations. But predicting the high demand is very important in real life, for better DBS management. Thus it appears that other factors such as POIs around the station and population are needed for better prediction.

Table 3 Top 10 subway stations with the highest DSB demand
Full size table

Figure 11 shows the pick-up demand prediction results for Guomao Station as an illustrative example. The pick-up demand peaks during weekdays are notably higher than those during weekends, indicating a higher usage of DSB for potential transfer trips on weekdays. The prediction model performs well for low-demand time slots, but is not able to accurately capture extremely high demand.

Fig. 11
figure 11

Guomao station pick-up demand prediction results

Full size image

Figure 12 presents the spatial results by fixing time slots for a weekday (May 19, 2021) evening peak of 7–8 pm and a weekend (May 23, 2021) evening peak of 6–7 pm across 236 subway stations, where the horizontal axis represents the station numbers. Similar to the temporal results shown previously, the method captures the trend well for most station predictions. However, for some high-demand stations during peak hours, the method underestimates the demand, indicating a noticeable prediction gap (Fig. 13).

Fig. 12
figure 12

May 19, 2021 19:00–20:00 drop-off demand prediction results

Full size image
Fig. 13
figure 13

May 23, 2021 18:00–19:00 drop-off demand prediction results

Full size image

Overall, it is postulated that other factors such as land use, POIs around subway stations, and population density may be necessary as explanatory factors for better peak demand prediction performance, particularly for subway stations with very high demand. The potential reason for this with regard to the method used in this study may be explained as follows:

  1. 1.

    Limited graph structure: AGCRN relies on graph structures to model relationships between entities. If the graph structure used in AGCRN does not capture the relevant relationships in high-demand situations, then the model may struggle to make accurate predictions.

  2. 2.

    Insufficient training data: The AGCRN may not have been trained on enough data from high-demand situations. As a result, the model may not have learned how to accurately predict demand when it exceeds a certain level.

5 Conclusion and Future Research

For better DSB and URT integration, this study processed 58 consecutive days of DSB usage data in Shenzhen from April 1, 2021, to May 28, 2021, with city-wide full-scale data (three operating companies’ data all included), defined the catchment area radius for the DSB demand around subway stations as 150 m, and analyzed the DSB demand around subway stations. The results show that there are substantial variations across URT stations in terms of supply and demand gap, especially in peak hours. Therefore, predicting the DSB demand in advance is critical for better DSB and URT management, and thus greater convenience for DSB and URT users.

Then this paper predicted the pick-up and drop-off demand at the 236 subway stations in 2021 using the most up-to-date deep learning method, namely AGCRN, and compared with four other deep learning methods: LSTM, STNN, DCRNN, and Graph WaveNet. AGCRN utilizes a graph convolutional network (GCN) to capture the spatial dependence of traffic flows and a recurrent neural network (RNN) to model the temporal patterns. The attention mechanism is incorporated to enhance the model's capability to attend to relevant spatial and temporal features, and the prediction results in this study show that AGCRN is better than the other four implemented deep learning methods in terms of MAE, MAPE, and RMSE. Thus, the AGCRN method has shown promising results in accurately predicting DSB demand in this study, making it a valuable tool for transportation planning and management.

This study provides very useful information in terms of subway station DSB demand prediction methods for large cities such as Shenzhen’s DSB-related policymakers and DSB operating companies, and might also be useful for DSB management and regulations for other cities. However, there are still limitations for this research, and the following are future research directions:

  1. 1.

    The available data pertain to the subway volume and bikes for the years 2018 and 2021, respectively. Additionally, the subway line data are from the end of 2022, which includes several recently opened subway lines. As a result, the transfer analysis was limited to the year 2018. However, if more data become available, a more comprehensive analysis of bike usage evolving through the years in relation to the subway can be conducted.

  2. 2.

    The catchment area method is an empirical estimation based on previous studies and real subway station geography in Shenzhen, while the catchment radius for all the stations uses the same number, which is 150 m at this time. However, a more accurate estimation of the station based on station size, land use, etc., is needed for future research.

  3. 3.

    The prediction methods in this study use the subway adjacency matrix together with historical data as data input, but do not include other data such as land use, URT station volume, weather, population, infrastructure, and construction, which are all considered or previously proven to influence the accuracy of shared bike ridership prediction. Thus, it is recommended that these data be added in future study to better capture the high-demand station situation to improve the graph representation of station characteristics and the correlation of these characteristics with station volume. The potential for increased accuracy is particularly noteworthy, as it pertains to the analysis of complex systems characterized by nonlinear relationships, and highlights the potential utility of graph-based deep learning methods in this domain. With the provision of additional data, particularly land use data, the method presented in this report is expected to exhibit enhanced predictive accuracy.