Bike-Sharing Travel Demand Forecasting via Travel Environment-Based Modeling

Abstract

This research aims to address the limited consideration given to non-motorized transport facilities in current studies on shared bike travel demand forecasting. This study is the first to propose a method that applies complete citywide non-motorized facility data to predict bike-sharing demand. This study employs a multiscale geographically weighted regression (MGWR) model to examine the effects of non-motorized transport facility conditions, quantity of intersections, and land use per unit area on riding demand at various spatial scales. The results of comparison experiments reveal that riding demand is substantially affected by non-motorized transport facilities and the quantity of intersections.

1. Introduction

In recent years, the proliferation of shared bicycles and their unrestricted parking in urban areas has led to inefficiencies in the utilization of pedestrian spaces and has posed challenges in managing and maintaining urban roads. The reason behind this is the conflict between riding demand and inadequate non-motorized transport facilities. Bike-sharing trips have high requirements for bicycle lane width, non-motorized transport facilities, and continuity. Suitable bicycle lane widths directly affect the service level of bike lanes [1]. Well-developed non-motorized transport facilities can boost the demand for riding [2]. However, the existing research has limitedly analyzed the relationship between the development of transportation facilities and the demand for bike-sharing trips, with only a small proportion of indicators focusing on transportation facilities. Furthermore, these studies have failed to conduct separate analyses on the construction of non-motorized transport facilities. In terms of research methodology, studies that consider transportation facilities have certain limitations. The two primary methodologies for investigating non-motorized transport facilities have limitations. First, surveys and questionnaires that inquire about bicycling participants’ perceptions of non-motorized transport facilities introduce a degree of subjectivity and may not provide entirely objective results [3]. Second, collecting data from selected road samples to evaluate small-scale road networks is data-limited and may not yield a comprehensive analysis [4].

To address these issues, this paper references research related to transportation facility evaluation and bike-sharing travel demand forecasting. By utilizing data on non-motorized transport facilities within extensive urban road networks, we devise a method for evaluating these facilities. Following this, we use the outcomes of the facility evaluation as independent variables and analyze their relationship with bike-sharing travel demand, taking into account factors such as intersection count, proximity to roads, and geographical points of interest. This approach aims to overcome the limitations of the existing research, particularly the scarcity of data on non-motorized transport facilities and the subjectivity of samples. The evaluation of non-motorized transport facilities incorporates both the analytic hierarchy process (AHP) and the entropy weight method. Subsequently, the entire city’s road network is divided into equally sized sub-regions using a grid method. This allows for the analysis of the relationship between non-motorized transport facility construction and bike-sharing trips in each region, ultimately enabling the construction of a demand prediction model.

This method effectively establishes a connection between the construction of non-motorized transport facilities in various regions of a large-scale urban road network and the corresponding number of shared bicycle trips in those regions. By doing so, it addresses the limitations present in the existing research, which include a small sample size of road non-motorized transport facilities and a lack of objectivity in data analysis.

2. Literature Review

2.1. Bike-Sharing Trip Demand Forecasting

In general, the factors that influence the demand for bicycle sharing are multifaceted. Thus far, the existing research has explored the impact of various factors, including departure time, weather conditions, location, and population distribution [5,6,7]. Additionally, different land use patterns also have the potential to influence cycling demand [8,9]. Non-motorized transport facilities emerge as a crucial factor in determining bicycle demand. Studies have revealed a complex nonlinear relationship between the demand for bike-sharing and factors such as the distance from the street to the city center, the proportion of non-motorized lanes, transportation facilities, population density, and Point of Interest (POI) density [10,11]. Advocates of bicycle-friendly cities propose the construction of robust bicycle road networks to foster increased bicycle usage [12]. Echoing this sentiment, a notable number of cyclists have expressed interest in cycling but expressed dissatisfaction with the current bicycle facilities [13,14].

For these influencing factors, commonly employed methods encompass machine learning and neural network models for forecasting bike-sharing demand. Sun Qipeng et al. [15] devised a forecasting model using a back-propagation (BP) neural network, which was based on a non-negative matrix factorization algorithm, to predict bike-sharing trip demand. Chen et al. [16] formulated a prediction model leveraging a BP neural network and selected historical bike-sharing usage data from the case area. Through programming and model resolution using Matlab software, Chen ascertained the predicted future bike-sharing usage volume in the designated area. Yan et al. [17] crafted a BP neural network model that integrates user tidal riding patterns and geographical spatial point-of-interest data to anticipate bike-sharing riding volume influenced by weather factors. Meanwhile, Lee et al. [18] developed six deep-learning models sharing the same architecture to forecast demand for a station-based bike-sharing service in Seoul, South Korea.

However, machine learning and neural network algorithms prove inadequate for all regions because the demand for bike-sharing trips commonly shows spatial variation. Furthermore, these algorithms are plagued by issues encompassing poor interpretability and substantial operational complexity [19,20].

2.2. Geographically Weighted Models

The geographically weighted regression (GWR) model has been increasingly employed to predict the demand for shared bicycle trips. The GWR-based model can consider spatial variations, thus allowing for exploring the spatial variations in a research object at a certain scale. Wu et al. [21] employed both a global regression model and a GWR model to analyze the global and local effects of the built environment on bicycle use. Ma et al. [22] built geographically and temporally weighted regression models to examine the influence of bike-share stations, socio-demographic factors, and land use factors on user demand for all bike-sharing systems.

Although the GWR model is capable of capturing the spatial variation in all influencing factors, it cannot reflect the spatial variation in the range of the single influencing factor, which affects the prediction accuracy [23]. Some researchers proposed the multiscale geographically weighted regression (MGWR) model based on the GWR model. This model can calculate the bandwidth of each influencing factor and has the advantage of analyzing the spatial variation in these factors and differences among adjacent areas [24]. Hao et al. [25] exploited the MGWR-based model to examine the impact of the density of intersections and points of interest on bike-sharing trip demand at various spatial scales. Guo et al. [26] utilized multivariable linear regression-based and MGWR-based models to quantify the spatial and temporal heterogeneity in the impact of the built environment on the vitality of different age groups in cities. Their study revealed that the MGWR-based model is capable of more precisely predicting riding demand in refined areas [27].

Based on the above summary, we utilized the MGWR model to analyze the influence of spatial variability in non-motorized transport facilities and land use on riding demand.

3. Data

3.1. Data Source

The bike-sharing data utilized in the present investigation were extracted from the operational data of Mobike (one of the largest bike-sharing companies in China) in Beijing. The data fields gathered include user ID, vehicle ID, date, longitude, and latitude of both the starting and ending points. The data pertaining to bicycle lane facilities in Shijingshan District originated from the relevant departments in the Shijingshan area, comprising a total of 3483 data points. The types of data and field information are presented in

Table 1. Bicycle lane facility data.

ID	Research Projects	Coverage	Field Information
1	Bicycle lane isolation belt	All roads	Marked isolation/green plants or footpaths isolation/mixed traffic
2	Bicycle lane Width	All roads	Bicycle lane width
3	Roadside parking spaces	All roads	Settings/not set
4	Shade cover	All roads	Facility belt with shade/isolation belt with shade/both/no shade cover
5	Color pavement setting rate	Intersection facilities	Qualified/not qualified
6	Bicycle lane damaged sign	Intersection facilities	No damage/medium damage/severe damage
7	Bicycle lane signal light setting rate	Intersection facilities	Settings/not set
8	Intersection data	All roads	Intersection coordinates

. The data on Beijing’s geographic points of interest (POI) were also obtained via Gaode’s open map API.

3.2. Data Pre-Processing

ArcGIS was utilized to generate a research area based on the bike lane data that covers the range of a one-kilometer radius around each road. The studied area was further divided into subareas of the same size, and a total of 1323 subareas were obtained, as illustrated in Figure 1.

4. Selection and Quantification of the Independent Variable

4.1. Evaluation of Non-Motorized Transport Facilities

Based on the data obtained for non-motorized transport facilities and urban road networks, this paper categorized and organized information and evaluated facility settings. An evaluation index was developed for bicycle lanes, which was primarily divided into three categories.

4.1.1. Evaluation of Roadway Facilities

From the perspective of the traffic capacity of bicycle lanes, the research conducted by Linjua et al. [28] showed that bicycle lanes with a width between 1.5 and 2.5 m increase the capacity with increasing width. When the lane width exceeds 2.5 m, there is little improvement in the bicycle lane capacity, and when the lane width is less than 1.5 m, the capacity is almost equivalent to mixed traffic. The classification of other bicycle lane infrastructures is performed based on their construction status. Figure 2 illustrates the differences in the widths of bike lanes.

This paper used the analytic hierarchy process to determine the weight of different facilities in influencing bicycle travel demand. The analytic hierarchy process (AHP) enabled us to decompose the problem into various constituent factors based on its nature and the overall goal to be achieved. By examining the interdependencies and hierarchical relationships among these factors, we organized them into a multi-layered analytical structure model. This process enabled us to determine the weights among different indicators. The scores were assigned based on four facility settings of various roads. The total score for each road was obtained by summing up the facility scores. After that, the average road score was obtained in each area by dividing the total road scores for that area by the number of roads within the unit area. The resulting score served as the independent variable for roadway facilities, denoted by the symbol X1. The scoring rules are presented in

Table 2. Roadway facility assignment score evaluation.

Lane Width	Score	Isolation Belt	Score	Roadside Parking Spaces	Score	Shade Cover	Score
<1.5 m	0	Mixed traffic	0	Not set	0	No shade cover	0
1.5–2 m	1	Scribing or guardrail	1	Settings	1	Single-sided cover	1
2–2.5 m	2	Greenery or footpaths	2	-	-	Both	2
>2.5 m	3	-	-	-	-	-	-

.

Then, we invited five experts in the field of transportation planning and management to construct a judgment matrix based on the relative importance of these four facilities. The judgment matrix is shown in

Table 3. Weight matrix table among different facilities.

	Lane Width	Isolation Belt	Roadside Parking Spaces	Shade Cover
Lane width	1	3	3	9
Isolation belt	0.333	1	1	5
Roadside parking spaces	0.333	1	1	5
Shade cover	0.111	0.2	0.2	1

.

The results of the analytic hierarchy process are shown in

Table 4. AHP hierarchical analysis results and consistency check.

Evaluation Indicators	Eigenvector	Weight Value (%)	The Maximum Eigenvalue	CI Value	RI Value	CR Value
Lane width	2.166	54.159	4.033	0.011	0.882	0.012
Isolation belt	0.822	20.553
Roadside parking spaces	0.822	20.553
Shade cover	0.189	4.736

. The results indicate that the weight of the lane width is 54.159%, the weight of the isolation belt is 20.553%, the weight of roadside parking spaces is 20.553%, and the weight of the shade cover is 4.736%. The calculation results of the analytic hierarchy process show that the maximum eigenvalue is 4.033. According to the RI table, the corresponding RI value is 0.882, thus CR = CI/RI = 0.012 < 0.1, passing the consistency test.

4.1.2. Evaluation of Intersection Facilities

This index was used to investigate the effect of constructing non-motorized transport facilities at road intersections on the riding demand in the study region. Because the data are not a complete sample, it only measures the construction of intersection facilities on high-grade roads in Shijingshan District, with a total of 19 roads. The setting rate of bicycle lane signal lights at intersections and the degree of damaged signs in bicycle lanes commonly reflect the quality of those infrastructures. As for colored pavement, it is not necessary to use it on a large scale. Instead, it can have the greatest impact near intersections with roads [28]. Figure 3 represents the pavement color settings and signage damage in the bike lanes.

The entropy weight method was employed to evaluate the placement of bicycle lane facilities, with the pavement setting rate, the signs qualified rate, and bicycle lane signal lights serving as positive indicators, whereas the moderate and severe damage rates of bicycle lane signs were the negative indicators. The entropy method is an objective weighting approach. Its principle lies in reflecting the magnitude of information based on the degree of variation among indicator values during the evaluation process, thereby determining the corresponding weights. Given the distinct differentiation between positive and negative indicators evident from the outset of data collection for this particular set of indicators, the adoption of the entropy method was deemed appropriate. The calculation results are provided in

Table 5. Example of an intersection facility evaluation.

Road Number	Color Pavement Setting Rate	Sign Qualified Rate	Signal Light Setting Rate	Sign Moderate Damaged Rate	Sign Severe Damaged Rate	Comprehensive Evaluation	Rank
1	0.849	0.833	0.397	0.131	0.036	0.772	8
2	1.000	0.303	0.875	0.424	0.273	0.557	16
3	0.250	0.900	0.750	0.1	0	0.760	9
4	1.000	0.742	0.129	0.129	0.100	0.672	10
5	1.000	0.500	0.500	0.000	1.000	0.772	7

.

Finally, the scores of intersection non-motorized transport facilities for 19 roads were obtained and matched with spatial data. Since intersections have a substantial impact on surrounding roads, this study established a 200 m buffer zone around each intersection. Then, the facility evaluation scores were applied to the unit areas within this range, which were affected by the evaluation results. The resulting score serves as the independent variable for intersection facilities, denoted by the symbol X2.

4.2. Number of Intersections and Distance from Adjacent Roads

Hao et al. [25] noted that the number of intersections in a given area affects the demand. Consequently, this study analyzed the quantity of intersections per unit area as an independent variable. To perform this analysis, ArcGIS was utilized to count the intersections of all roads in the road network data. This count served as the independent variable for the quantity of intersections, denoted by the symbol X3. Since the road network data do not include branch roads, this study introduced the “adjacent roads distance” indicator to examine the relationship between bike-sharing trip demand and the distance between main roads. The distance of adjacent roads measures the distance of each area unit to the nearest main road. When the main road data were available in an area, the distance between the area and the road would be 0 m. However, if there were no major road data in an area, the distance from the area to the nearest major road was calculated. The resulting value served as the independent variable for the adjacent road distance, denoted by the symbol X4.

4.3. Evaluation of Land Use

This section evaluation selected 12 types of land use, and the geographic points of interest were employed in the Shijingshan area. Then, these were appropriately matched with spatial data and the counted number of each type of land use in various areas to serve as explanatory variables, represented by symbols X5 through X16, which are presented in

Table 6. Summary of various variables.

Factors	Variable Name	Variable Symbols	Description
Dependent variable	Departure volume	Y	Unit regional departure volume
Traveling environment	Roadway facilities	X1	Comprehensive evaluation
	Intersection facilities	X2	Comprehensive evaluation
	Quantity of intersections	X3	Quantity of unit area intersections
	Adjacent road distance	X4	Distance from nearest road (m)
Land use	Company	X5	Quantity in the unit area
	Bus and subway station	X6
	Residential area	X7
	Research and education	X8
	Tourist attractions	X9
	Commercial areas	X10
	Leisure sports	X11
	Health care	X12
	Public restrooms	X13
	Government	X14
	Life services	X15
	Financial sector	X16

.

5. Methodology

5.1. Development of the GWR-Based Model

The GWR model, as a locally based spatial model, addresses the following spatial issues: spatial autocorrelation and spatial heterogeneity. This is accomplished by setting various spatial weights in each local area and incorporating varying estimated parameters in each location. The construction of the GWR model essentially originated from the following relation:

$$y_i={\beta }_{i0}+{\sum }_{k=1}^m{\beta }_k(u_i,v_i)x_{ik}+{\epsilon }_i{y}_i={\beta }_{i0}+{\sum }_{k=1}^m{\beta }_k(u_i,v_i)x_{ik}+{\epsilon }_i\hspace{1em}i=1,2,\dots ,n$$

(1)

where $y_i$ denotes the riding demand at the i-th region, ${\beta }_{i0}$ represents the regression intercept of the i-th region, $x_{ik}$ is the k-th riding demand influence factor in the i-th region, ${\beta }_k\left(\right)$ signifies the regression coefficient of the k-th riding demand influence factor in the i-th region, and ${\epsilon }_i$ is the random error term of riding demand at the i-th region. In matrix form, the local coefficient of the i-th region, ${\beta }_k\left(u_i,v_i\right),$ is estimated by:

$$\stackrel{\wedge }{{\beta }_k}(u_i,v_i)={\left[{\mathrm{X}}^T\mathrm{W}(u_i,v_i)\mathrm{X}\right]}^{-1}{\mathrm{X}}^T\mathrm{W}(u_i,v_i)y\stackrel{\wedge }{{\beta }_k}(u_i,v_i)={\left[{\mathrm{X}}^T\mathrm{W}(u_i,v_i)\mathrm{X}\right]}^{-1}{\mathrm{X}}^T\mathrm{W}(u_i,v_i)\mathrm{y}$$

(2)

where X denotes the n × k matrix of riding demand influencing factors, $\mathrm{W}\left(u_i,v_i\right)=diag[w_{i,1},\cdots ,w_{i,n}]$ denotes the weight matrix of n × n, which weights the riding demand influencing factors on each observation region based on the centroid distance between each observation region and regression in the i-th region, is the coefficient vector of k × 1, and y represents the vector of riding demand quantity observations of k × 1.

The kernel function was utilized to calculate the weight of the distance between the observed region and the calibration region. The commonly used kernel functions include Gaussian, exponential, and quadratic functions. As the Gaussian function is relatively smoother and closer to reality, we used a Gaussian kernel function for distance weight calculation, as given by Equation (3):

$$w_{ij}=exp(-{\displaystyle \frac{1}{2}}({\displaystyle \frac{d_{ij}}{b}}{)}^2)$$

(3)

where $d_{ij}$ denotes the centroid distance between the j-th observation region and the i-th regression region, the so-called kernel bandwidth, and $w_{ij}$ represents the weight that makes up the weight matrix $\mathrm{W}(u_i,v_i)$.

The kernel bandwidth is the key control parameter of GWR-based regression, and the bandwidth with the smallest AICc was chosen as the optimal kernel bandwidth by the cross-validation method in this paper. The formula of AICc is provided by the following relationship:

$$\mathrm{A}\mathrm{I}\mathrm{C}\mathrm{c}=2n\ln (\stackrel{\wedge }{\sigma })+n\ln (2\pi )+n{\displaystyle \frac{n+\mathrm{t}\mathrm{r}\left(\mathbf{S}\right)}{n-2-\mathrm{t}\mathrm{r}\left(\mathbf{S}\right)}}$$

(4)

where n is the number of samples, $\stackrel{\wedge }{\sigma }$ is the standard deviation of the error term estimates, and $\mathrm{t}\mathrm{r}\left(\mathbf{S}\right)$ is the trace of the S-matrix of the GWR, which is a function of the bandwidth.

5.2. Development of the MGWR-Based Model

The GWR model aims to examine the potential spatial instability in relationships and provides a measure of the spatial scale at which processes operate optimally by determining bandwidth. The classical GWR-based model assumes that all modeled processes operate at the same spatial scale. However, the MGWR-based model overcomes this limitation by assuming that each variable has a unique spatial bandwidth. This enables the model to remove the constraints that all explanatory variable relationships in the GWR-based model are different at the same spatial scale. As a result, the error bias in parameter estimation can be rationally reduced, and the problem of multicollinearity is appropriately reduced [24]. The essential equation of the MGWR-based model is presented by:

$$y_i={\beta }_{bw0}+{\sum }_{k=1}^m{\beta }_{bwk}(u_i,v_i)x_{ik}+{\epsilon }_i$$

(5)

where ${\beta }_{\mathrm{bw}0}$ denotes the intercept of regression of the i-th region, ${\beta }_{\mathrm{bwk}}(u_i,v_i)$ is the coefficient of the influence of the k-th riding demand influence factor on the bike-sharing trip demand at the regression of the i-th region, and $\mathit{bwk}$ denotes the bandwidth of the k-th influence riding demand factor on the bike-sharing trip demand.

The ${\beta }_{\mathrm{bwk}}(u_i,v_i)$ estimation method of the MGWR-based model is similar to that of the GWR-based model. The weight calculation is calculated by the following relation:

$$w_{ij}=\exp (-{\displaystyle \frac{1}{2}}({\displaystyle \frac{d_{ij}}{bwk}}{)}^2)$$

(6)

6. Results and Analysis

6.1. Comparison of the Demand Forecast Models

We used the ordinary least squares regression (OLS) model, GWR model, and MGWR model to complete regression analysis and compare their obtained results. The comparison results are provided in

Table 7. Comparison of the model parameters.

Model	R2	Adjusted R2	RSS	AICc	Koenker (BP)
OLS	0.406	0.385	785.487	3101.286	0.000000 *
GWR	0.509	0.489	645.577	2995.588	-
MGWR	0.641	0.601	474.542	2794.932	-

Note: * represents significant results.

. The Koenker (BP) statistics serve as a test method to ascertain whether the variables of a model exhibit an identical relationship with the dependent variable in both geospatial and data spaces. When the test results are statistically significant, it indicates that the regression model is more suitable for application in geographically weighted regression analysis.

It can be seen that the fitting effect of the MGWR model is higher than that of the other two models. Figure 4 illustrates a comparison of the local R² between the GWR model and the MGWR model. The red ellipses mark the areas with significant differences in R².

6.2. Bandwidth Analysis

Table 8. Bandwidth results of the GWR and MGWR models.

Variable Name	GWR-Based Bandwidth	MGWR-Based Bandwidth	Variable Name	GWR-Based Bandwidth	MGWR-Based Bandwidth
Riding environment	474	1321	Tourist attractions	474	1321
Intersection facilities	474	59	Commercial areas	474	284
Quantity of intersections	474	57	Leisure sports	474	1190
Adjacent road distance	474	1321	Health care	474	940
Company	474	416	Public restrooms	474	159
Bus and subway station	474	268	Government	474	193
Residential area	474	450	Life services	474	1321
Research and education	474	1314	Financial sector	474	746

presents the bandwidth values for both the GWR and MGWR models. In the MGWR model, the scales of the variables differ, whereas the GWR model can only provide an average of the action scales for each variable.

Table 8 shows that most of the explanatory variables exhibit spatial variations in their impact on the riding demand, with only a few variables having consistent effects across the entire study area. For instance, the bandwidth of the roadway facilities, trip attractions, and other variables are high, indicating their impact on the entire region. However, the values of intersection facility bandwidth, quantity of intersections, and other variables are low, which indicates the limited impact range. These indicators only affect the riding demand in the adjacent area, but do not affect the riding demand in the nearby area.

6.3. Analysis of the Factors Influencing Demand

Table 9. Regression results of the OLS and MGWR models.

Variable Type	Variable Name	OLS			MGWR
		Regression Coefficient	p-Value	Significance	p-Value	Significance
Traveling environment	Roadway facilities	0.061	0	***	0.035	*
	Intersection facilities	0.027	0.001	**	0.001	**
	Quantity of intersections	0.044	0	***	0.001	**
	Adjacent road distance	0.005	0.583	N/A	0.033	*
Functional area	Company	0.011	0.75	N/A	0.012	*
	Bus and subway station	0.087	0	***	0.005	**
	Residential area	0.094	0	***	0.011	*
	Research and education	0.083	0.015	*	0.042	*
	Tourist attractions	0.06	0.303	N/A	0.037	*
	Commercial areas	0.243	0	***	0.008	**
	Leisure sports	−0.005	0.86	N/A	0.041	*
	Health care	0.036	0.116	N/A	0.027	*
	Public restrooms	−0.034	0.029	*	0.003	**
	Government	0.072	0.012	*	0.005	**
	Life services	−0.045	0.267	N/A	0.043	*
	Financial sector	0.007	0.832	N/A	0.023	*

Note: ***, **, and * represent significance levels of 1%, 5%, and 10% respectively, N/A represents not significant.

presents the fitting results of both the OLS and MGWR models. It can be seen that adjacent road distance, company, tourist attractions, leisure sports, health care, life services, and financial units are not significant in the OLS-based regression results, but they are remarkable in the MGWR-based model. The results show that these variables are not irrelevant to the demand for riding, but there is a large spatial variation in the degree of impact, so the demand forecast needs to be refined to a smaller spatial area.

Table 10. Summary of the MGWR-based regression coefficients.

Variable Type	Variable Name	Mean	Min	Max	Median
Traveling environment	Roadway facilities	0.095	0.081	0.107	0.097
	Intersection facilities	0.119	−0.313	0.967	0.064
	Quantity of intersections	0.088	−0.279	1.219	0.032
	Adjacent road distance	0.006	0.002	0.013	0.005
Functional area	Company	0.048	−0.088	0.233	0.022
	Bus and subway station	0.093	−0.035	0.362	0.026
	Residential area	0.085	0.019	0.184	0.066
	Research and education	0.047	0.04	0.058	0.046
	Tourist attractions	0.047	0.016	0.093	0.05
	Commercial areas	0.255	0.119	0.428	0.251
	Leisure sports	0.037	0.013	0.058	0.037
	Health care	0.053	0.009	0.083	0.063
	Public restrooms	−0.031	−0.286	0.313	−0.014
	Government	0.031	−0.236	0.405	0.003
	Life services	−0.019	−0.021	−0.015	−0.02
	Financial units	0.068	−0.024	0.134	0.071

presents the statistical description of the regression coefficients for each variable in the MGWR model, while Figure 5 illustrates the spatial distribution of these coefficients. In Figure 5, the darker areas represent the regression coefficients of the influential factors with significant influence, whereas the gray squares denote the regression coefficients of the influential factors with insignificant influence.

Table 10 shows that roadway facilities significantly affect the demand for bike-sharing and have a positive effect, meaning that better roadway facilities lead to greater demand for bike-sharing in the same region. Figure 5a demonstrates that the roadway facilities coefficient is in the range of 0.081–0.107, with a gradual decrease in the influence range from west to east in terms of spatial distribution, with the strongest impact intensity in the west. There are two main reasons for this. First, the northwest area of Shijingshan District is adjacent to the West Mountain, which has a total area of 5970 hectares and is 300–400 m high. This makes the construction of non-motorized transport facilities in this area more challenging than in the east and southeast, which are flat and have better road facilities. Second, the southeast region is located within Beijing City and is close to Haidian District, which has a relatively complete non-motorized transport facility.

Intersection facilities have a significant positive effect on the riding demand, as presented in Figure 5b, with a coefficient effect in the interval of [−0.313, −0.967]. This influence is mostly concentrated in the eastern and central areas of Shijingshan District. The reason for this phenomenon could be attributed to both geographical factors and urban development plans. The concentration of intersection facilities in the eastern and central areas of Shijingshan District leads to a more pronounced impact of these facilities in these areas.

As shown in Figure 5c, the quantity of intersections also has a positive effect on the riding demand, with a coefficient effect ranging in the interval of [−0.279, 1.219]. This impact is more noticeable in the middle and southeast of the region and less in the northwest. A possible explanation for this phenomenon is that because of the influence of the mountains, the overall development of the eastern region is better than that of the western region.

According to land use, commercial areas, bus and subway stations, and residential areas exhibit a significant positive impact on shared-bike demand with coefficients ranging in the intervals of [0.119–0.428], [−0.035–0.362], and [0.019–0.184], respectively. Specifically, bus and subway stations essentially affect the riding demand in the central parts, commercial areas have more influence on the eastern and central parts, and residential areas mainly affect the riding demand in the central parts of Shijingshan District. This is probably due to the geographical distribution of these functional areas including research and education, tourist attractions, leisure sports, health care, public restrooms, government, life services, and financial units. Actually, the coefficients of these factors and their spatial distribution do not follow a clear rule. The reason for this phenomenon may be that some small-scale geographic point-of-interest data were not filtered during the data screening process, resulting in errors that affect the results.

7. Conclusions

This study, for the first time, forecasts bike-sharing demand based on non-motorized transport facility construction across the entire urban district road network. We used the analytic hierarchy process (AHP) and entropy weight method to evaluate the traffic facilities on each road in Shijingshan District. We divided the city network into 1323 zones to analyze traffic development. Then, we employed the MGWR-based model to forecast bike-sharing demand from the perspective of the trip environment and infrastructure. The main conclusions are summarized as follows:

(1)

The model in this paper predicts bicycle travel demand based on the construction of non-motorized facilities in either the whole area or local areas. This provides urban managers with important information for optimizing non-motorized facilities.

(2)

The MGWR model is able to reflect spatial variations and differences in the scale and degree of influence of each factor, enabling a more accurate prediction of riding demand in refined areas. The goodness of fit of the MGWR model is noticeably higher than the other two models.

(3)

The more comprehensive the non-motorized transport facilities system on urban roads, the more conducive it is in promoting the growth of bike-sharing travel demand. However, the impact coefficients of non-motorized transport facilities improvements vary across different regions. Relevant departments or agencies can enhance bike-sharing travel demand by improving facilities in key areas. Generally speaking, compared with other types of areas, improving the facilities surrounding residential areas, commercial districts, and public transportation stations can achieve a more significant increase in bike-sharing travel demand.

(4)

Based on the experimental results, non-motorized transport facilities in intersection areas exhibit higher significance and impact coefficients in the model compared with those in roadway segments. This underscores the fact that improving non-motorized transport facilities within intersection zones has a more pronounced effect on promoting shared bike utilization. Consequently, when enhancing non-motorized transport facilities, priority should be given to upgrading those in intersection areas. As for roadway non-motorized transport facilities, lane width stands as a paramount indicator, and relevant authorities should adjust the widths of bicycle lanes in different regions to accommodate the unique characteristics of each area.