Hindawi

Mathematical Problems in Engineering

Volume 2018, Article ID 8028714, 8 pages
https://doi.org/10.1155/2018/8028714

Research Article
A Markov Chain Based Demand Prediction Model for
Stations in Bike Sharing Systems

Yajun Zhou, Lilei Wang , Rong Zhong, and Yulong Tan

School of Transportation and Logistics, Southwest Jiaotong University, No. 111 North Second Ring Road, Chengdu,
Sichuan 610031, China
Correspondence should be addressed to Lilei Wang; wanglilei [email protected]

Received 31 May 2017; Revised 10 November 2017; Accepted 3 December 2017; Published 3 January 2018

Academic Editor: Hakim Naceur

Copyright © 2018 Yajun Zhou et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Accurate transfer demand prediction at bike stations is the key to develop balancing solutions to address the overutilization or
underutilization problem often occurring in bike sharing system. At the same time, station transfer demand prediction is helpful to
bike station layout and optimization of the number of public bikes within the station. Traditional traffic demand prediction methods,
such as gravity model, cannot be easily adapted to the problem of forecasting bike station transfer demand due to the difficulty in
defining impedance and distinct characteristics of bike stations (Xu et al. 2013). Therefore, this paper proposes a prediction method
based on Markov chain model. The proposed model is evaluated based on field data collected from Zhongshan City bike sharing
system. The daily production and attraction of stations are forecasted. The experimental results show that the model of this paper
performs higher forecasting accuracy and better generalization ability.

1. Introduction A full understanding of demand is a crucial step to

improve the prediction accuracy. Different bike sharing
Bike sharing systems are in place in many cities in the world systems may be divergent; nevertheless, significant influence
and are an increasingly important support for multimodal factors are the same, such as lanes, population, economic and
transport systems [1, 2]. The imbalance between production social conditions, festival, workday, weather, and land use [5–
and attraction from stations is one of the greatest problems 8]. Different factors in different time periods cause different
in practical system operation at present [3], thus making influence degree. Generally, demand influence factors like
users unable to rent or return bikes for a time, hindering lanes, population, and economic level can be seen as constant
systems to normally operate, and limiting further promotion. in a day or smaller time unit. Compared to the overall
Currently, the main solution to solve the imbalance of system usage of bike sharing system, the station’s demand has its
is proceeding to dispatch bikes inefficiently among stations. own particularity. Apart from considering factors influencing
Accurate demand prediction can offer a beneficial guide overall demand, we also need to keep an eye on station’s
for managers to plan and design purposefully and thus feature. Public bikes move among stations; the amounts of
would help to solve the imbalance between production and production and attraction in a station are closely related to
attraction from stations. other stations. An inevitable constraint exists among stations’
Bike sharing system consists of bikes, roads, and fixed traffic volume rarely appearing in other researches at present
stations, whose demand has clear deference with motor [9].
vehicle and private bike. The demand forecasting model of Easily operated and effective regression model is the
motor vehicle and private bike can hardly be adapted to bike main method to forecast the usage of bike sharing system
sharing system [4]. Therefore, it is increasingly important at present, considering important demand influence factors,
to explore reasonable demand forecasting model which is such as population, weather, workday, land use, and envi-
matching sharing bike’s feature, but there are few studies ronment [10, 11]. Importantly, using regression model to
regarding its demand and demand prediction now. forecast demand, comprehensive understanding of influence
 2629, 2018, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2018/8028714 by Turkey Cochrane Evidence Aid, Wiley Online Library on [17/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 Mathematical Problems in Engineering

factors is the key to improve the prediction accuracy. Single among stations. Therefore, we put forward a hybrid model
regression model can hardly adapt the demand prediction based on Markov chain and have a test to inspect it.
of bike sharing systems, so a variety of multiple regression
models to forecast, respectively, can help to obtain optimal 2. Methodology
results [12]. Apart from regression models, other methods
including Fuzzy Inference Mechanism [13] and hybrid model 2.1. Discrete-Time Markov Chain Model. Markov process is
[14] also have good application to forecast the whole demand widely used in modeling the dynamics of stochastic systems
of system. It is worth noting that the hybrid model represents and the state transitions of complex stochastic systems. A
an important development direction of bike sharing demand Markov process {𝑋(𝑡), 𝑡 ∈ 𝑇} is a stochastic process with
forecasting methods, which may eliminate inherent defects of the property that, given the system state at time 𝑡, 𝑋(𝑡), for
single model and withhold advantages of various models. a time 𝑠 > 𝑡, the system state 𝑋(𝑠) is not influenced by the
Above is the summary of the whole demand forecast- system states, 𝑋(𝑢) for 𝑢 < 𝑡, that is, prior to the time 𝑡. 𝑃 is a
ing of system, which can provide invaluable references to transfer matrix; the matrix elements are not negative, and the
station-level demand prediction. Stations are the basic unit sum of all the various elements is equal to 1, expressed in the
of bike sharing system; station-level demand impacts the probability of each element. The element in the matrix is the
system’s planning, design, and operation directly. However, probability that the bicycle will be retained, acquired, or lost.
few studies focus on the station-level demand forecasting. 𝑃(𝑛) represents the 𝑛 step transfer matrix.
There exists an obvious difference between the whole demand
of system and station-level demand, due to the station’s 𝑝11 𝑝12 ⋅ ⋅ ⋅ 𝑝1𝑗
[ ]
particularity and constraint among stations. So the method [𝑝21 𝑝22 ⋅ ⋅ ⋅ 𝑝2𝑗 ]
[ ]
of the whole demand prediction cannot suit well station-level 𝑃=[
[ .. .. .. .. ]
] (0 ≤ 𝑝𝑖𝑗 ≤ 1) . (1)
demand forecasting; more efforts should be taken to find the [ . . . . ]
[ ]
reasonable prediction methods of station-level demand.
Traditional station-level demand forecasting methods are [ 𝑝𝑖1 𝑝 𝑖2 ⋅ ⋅ ⋅ 𝑝 𝑖𝑗 ]
still mainly based on regression models, which fully consider In other words, the probability of any particular future
influence factors [15]. However, few station-level demand behavior of the process, if its current state is known, will not
predictions used by regression model can consider the traffic be altered by any additional knowledge concerning its past
usage constraints among stations. Some scholars amend this behavior. A discrete-time Markov chain is a Markov process
problem, for instance, Rixey R added station significance to whose state space is a finite or countable set with time interval
the set of variables and built up a regression model to forecast indexes as 𝑡 = (0, 1, 2, . . .). In formal terms, the Markov
station’s demand of three American bike sharing systems [16]. property is that
However, even if most factors have been considered, there
is still inherent flaw in using regression model to forecast 𝑝 {𝑋 (𝑡𝑛 ) = 𝑥𝑛 | 𝑋 (𝑡1 ) = 𝑥1 , 𝑋 (𝑡2 ) = 𝑥2 , . . . , 𝑋 (𝑡𝑛−1 )
station’s demand: the traditional method takes traffic zone as (2)
forecast unit, but one traffic zone may contain several stations, = 𝑥𝑛−1 } = 𝑝 {𝑋 (𝑡𝑛 ) = 𝑥𝑛 | 𝑋 (𝑡𝑛−1 ) = 𝑥𝑛−1 } .
which generates obstacles to forecast every station’s demand The probability of 𝑋𝑛+1 being in state 𝑗 given that 𝑋𝑛 is
precisely. To solve this problem, some scholars establish in state 𝑖 is called the one-step transition probability and is
ARIMA [17] and modified ARIMA [18] to forecast station’s denoted by
demand; their results show these new methods have a good
predictive accuracy. At the same time, Bayesian network 𝑃 (𝑋𝑛+1 = 𝑥𝑗 | 𝑋𝑛 ) . (3)
has also been used to forecast the station’s demand. These
new methods make contributions to the station’s demand Multistep transition probability can be calculated accord-
prediction, but their effectiveness needs more tests to be ing to one-step transition probability and the Markov prop-
explained [19]. erty as follows:
Please note that the lack of standardized evaluation
procedure (data, duration, error metric, etc.) forbids doing a 𝑃 (𝑋𝑛+2 | 𝑋𝑛 ) = ∫ 𝑃 (𝑋𝑛+2 , 𝑋𝑛+1 | 𝑋𝑛 ) 𝑑𝑋𝑛+1
fair comparison between them. Table 1 summarizes the main (4)
studies in bike sharing prediction, helping readers to grasp
= ∫ 𝑃 (𝑋𝑛+2 | 𝑋𝑛+1 ) 𝑃 (𝑋𝑛+1 | 𝑋𝑛 ) 𝑑𝑋𝑛+1 .
the research status.
Through above analysis, we can conclude that regression
If the transition matrix 𝑃 is irreducible and a periodic, the
model is an ideal method to forecast the whole demand of
𝑘-step transition matrix converges to a stationary distribution
bike sharing system. However, regression model is not very
𝜋 with each column different; that is,
suitable for station-level demand forecasting. In addition, it is
uncertain whether the other methods, like ARMIA, Bayesian lim 𝑝𝑘 = 𝜋. (5)
𝑘→∞
network, and so on, can work well for different bike sharing
systems. According to Markov chain properties in finite state
The main purpose of this paper is to build up a reasonable space, stationary distribution 𝜋 can be calculated by (5):
and efficient station’s demand forecasting model, which con-
siders most significant influence factors and the constraints 𝜋𝑝 = 𝜋. (6)
 Table 1: The summary of public bike demand forecasting studies.
Publishing
Authors City Timespan Object Methods Error metric
Mathematical Problems in Engineering

year
Relative Absolute Error
(29.98%), Root
Fanaee-T and Gama [11] Washington DC 2 years Weka’s regressors 2014
Relative Squared Error
(39.27%)
Borgnat et al. [10] Lyon 2 years + 8 months Linear regression Mean Relative Error: 12% 2011
The system The best performing one
Giot and Cherrier [12] Washington DC 2 years demand Various regression 2014
clearly beats the baselines
A hybrid model: clustering, Error
Xu et al. [14] Hangzhou 6 months 2013
SVM rate (3.57%)
The min Relative Absolute
Salaken et al. [13] Washington DC 2 years Fuzzy Inference Mechanism 2015
Error: 75.41%
Cagliero et al. [19] New York 13 months Bayesian Precision (79.1%) 2016
Froehlich et al. [20] Barcelona 13 weeks Bayesian network Relative Absolute Error: 0.08 2009
Three operational US 11 months + 6 months + 6
Rixey [16] Linear regression Adjusted 𝑅2 : 0.802 2013
systems months
2
Singhvi et al. [15] New York 1 month Station-level Regression Adjusted 𝑅 : 0.745 2015
Kaltenbrunner et al. [17] Barcelona 7 weeks demand ARIMA Mean Absolute Error: 1.39 2010
Root Mean Square Error
Yoon et al. [18] Dublin 27 days Modified ARIMA 5 min: 0.91, 60 2014
min: 3.47
Fournier et al. [21] Ottawa 1 year Regression Error rate (20.8%) 2017
3

2629, 2018, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2018/8028714 by Turkey Cochrane Evidence Aid, Wiley Online Library on [17/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
 2629, 2018, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2018/8028714 by Turkey Cochrane Evidence Aid, Wiley Online Library on [17/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 Mathematical Problems in Engineering

Markov chain is a special case of Markov process defined where 𝜋 represents steady-state probability vector for bike
as follows. rentals, 𝜋∗ represents steady-state probability vector for bike
(1) Assume {𝑋(𝑡), 𝑡 ∈ 𝑇} is a random process, if the state returns (attraction), and
of {𝑋(𝑡), 𝑡 ∈ 𝑇} is known at 𝑡0 and is irrelevant to the states of
{𝑋(𝑡), 𝑡 ∈ 𝑇} at 𝑡 > 𝑡0 , then we say {𝑋(𝑡), 𝑡 ∈ 𝑇} has Markov 𝜋𝑗 ≥ 0,
property.
𝜋𝑗∗ ≥ 0,
(2) Assume 𝑆 is the state space of {𝑋(𝑡), 𝑡 ∈ 𝑇}, for 𝑛 ⩾ 2
and 𝑡1 < 𝑡2 < ⋅ ⋅ ⋅ < 𝑡𝑛 ∈ 𝑇, 𝑋(𝑡𝑖 ) = 𝑥𝑖 , 𝑥𝑖 ∈ 𝑆, 𝑖 = 1, 2, . . . , 𝑛−1. 𝑛
The conditional probability of a future state with respect to a ∑𝑝𝑖𝑗 = 1, (11)
past state can be written as follows: 𝑖=1

𝑛
𝑃 {𝑋 (𝑡𝑛 ) ≤ 𝑥𝑛 | 𝑋 (𝑡1 ) = 𝑥1 , 𝑋 (𝑡2 ) = 𝑥2 , . . . , 𝑋 (𝑡𝑛−1 )
(7) ∑𝑞𝑖𝑗 = 1.
𝑖=1
= 𝑥𝑛−1 } = 𝑃 {𝑋 (𝑡𝑛 ) ≤ 𝑥𝑛 | 𝑋 (𝑡𝑛−1 ) = 𝑥𝑛−1 } .
The steady-state probability vector of bike rental and
Then, {𝑋(𝑡), 𝑡 ∈ 𝑇} is Markov process. The proposed demand
return can be calculated by (9) using 𝑃 and 𝑄.
forecasting model will be based on the Markov chain process
by considering the bike transfer matrix at bike stations as the Step 6 (predict the production and attraction of stations).
system state.
𝑟=𝑆∗𝜋
2.2. The Application in Bike Demand Estimation. The pro- (12)
posed Markov chain model for bike demand estimation will ℎ = 𝑆 ∗ 𝜋∗ ,
predict the probability of rental and returns at each station.
The probability will then be converted into the actual bike where 𝑆 represents the whole traffic flow of system we
transfer numbers based on the total travel numbers of rental have predicted in Step 1, 𝑟 represents the vector of stations’
bikes. The detailed algorithm is as follows. predictive production, and ℎ represents the vector of stations’
predictive attraction.
Step 1 (predict the whole traffic flow of bike sharing system).
This paper will select a variety of effective models to forecast 3. Empirical Study: A Case Study
the whole traffic flow of system, like regression model, neural from Zhongshan City
network, SVM, hybrid models, and so on. According to the
prediction accuracy of each model, we choose the best one as 3.1. Zhongshan Bike Sharing System. Zhongshan is a me-
the final prediction. dium-sized city in China, with a population of approximately
600,000 and the urban area of about 170 square kilometers.
Step 2 (construct the bike sharing transfer matrix between The city’s bike sharing system was first launched in 2011. Up to
each station). An 𝑛 × 𝑛 matrix is first built to store the now, it has developed with an inventory of about 7000 bikes,
bike sharing transfer data where 𝑛 represents the number which spreads over 167 stations in the urban area. The bike
of stations. The matrix element 𝑎𝑖𝑗 represents the number of renting requires users to pay by a fare card that is linked to a
bikes rented from the station i and returned to the station 𝑗. user account. However, the bike usage is free in the first hour
of a day. The mode share of bikes in Zhongshan is currently
Step 3 (construct the transition probability matrix 𝑃 for bike at 20%, among which shared bikes occupy 5%. The utilization
rentals (production)). Assume 𝑃𝑖𝑗 is an element of the matrix ratio of Zhongshan bike sharing system has reached 97%
𝑃; then and average daily usage frequency is 11,500 times. Being
𝑎𝑖𝑗 welcomed by users and operated stability of Zhongshan bike
𝑃𝑖𝑗 = 𝑛 . (8) sharing system is the important reason why we select it as
∑𝑖=1 𝑎𝑖𝑗
predictive object.
Step 4 (construct the transition probability matrix 𝑄 for bike In general, bike sharing systems will experience an unsta-
returns (attraction)). Assume 𝑞𝑖𝑗 is an element of the matrix ble period and a stable period from construction planning
𝑄: to normal operations. Zhongshan bike sharing system was
put into operation in 2011. In order to ensure the stability
𝑏𝑖𝑗 of the system within the predicted time, this paper selects
𝑞𝑖𝑗 = 𝑛 . (9) data in 2013 as the training data and forecasts the demand in
∑𝑖=1 𝑏𝑖𝑗
2014. Excluding the invalid data and the usage information of
Step 5 (calculate of the steady-state probability vector). The administrators, we get 5,645,070 times travelling data, includ-
model uses the stationary distribution property of Markov ing 365 days’ information in 2013 and 224 days’ information
chain model; namely, in 2014. Each data includes station number, time, user card
number, bike number, fee, and other information of renting
𝜋𝑃 = 𝜋 and returning shared bikes. Some useful information can be
(10) obtained by mining these data, like the trip matrix, spatial and
𝜋∗ 𝑄 = 𝜋∗ , temporal information of users, and so on.
 2629, 2018, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2018/8028714 by Turkey Cochrane Evidence Aid, Wiley Online Library on [17/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Mathematical Problems in Engineering 5

Table 2: Error rate comparison of these models.

Methods MLP SVM Linear regression

The average
0.26 0.174 0.115
relative error
(i) c : 1
The train error: (ii) Kernel: (i) R: 0.821
Description
0.149 RBF (ii) DW: 0.961
(iii) 𝑔: 1/k

model has the lowest relative error among these three meth-
ods. So, the predictive result of regression model is taken as
Figure 1: The distribution of bike sharing stations’ cumulative usage the ultimate one.
in 2013 in Zhongshan. In the multilayer perceptron neural network, three layers
of network (1 hidden layer) are set up, the number of nodes is
taken as few hidden nodes as possible, the input layer 7 nodes
are selected, the output layer is 1 node, and the hidden layer
is 5 nodes. The initial weight value of a random generator
This paper selects 167 public bike stations in Zhongshan
is designed to generate a random number of −0.5∼+0.5.
urban area as the research object. For the convenience of
The minimum training rate is 0.9. The dynamic parameter
readers to know the distribution of stations’ location and their
selection value is 0.7. Allow the error to be generally 0.001∼
traffic volumes intuitively, we did a survey of 167 stations’
0.00001. The number of iterations is 1000 times. Sigmoid
traffic volumes in 2014, as shown in Figure 1. The dot in
parameter is 0.9.
picture indicates the actual location of the station, and the size
In the SVM model parameters with RBF kernel, the
of the circle reflects the station’s traffic volume. The larger the
parameters of the initial options for “𝑠, 𝑡, 𝑐, 𝑔,” where 𝑠 =
circle is, the greater the value of traffic volume is.
0 means choosing the support vector machine (SVM), 𝑐-
Figure 1 shows that the distribution of shared bike type SVC, 𝑡 = 0 means choosing RBF kernel function, the
stations’ traffic volume is uneven; particularly the traffic parameter 𝑐 means punishment coefficient (defaults to 1), in
volume of station which locates outside city is generally low. general, the value of penalty coefficient must be moderate,
Such distribution situation has caused great difficulties for the excessively low parameter cannot meet the requirement of the
station’ demand forecasting. toolkit (e.g., when 𝑐 = 0, stop running the toolkit), excessive
penalty parameter 𝑐 causes overlearning, and parameter 𝑔
3.2. The Prediction of Bike Sharing System’s Demand. The represents the gamma function value in the kernel function
prediction of the whole demand is the first step of station’s (the default value is 1/𝑘, and 𝑘 is the number of feature
demand forecasting; the precision of this has a major influ- attributes of the sample data, and 7 feature vectors are selected
ence on the final result. The prediction period of this article here).
is shown by day; one prediction is done per day. So, when Usually initial parameters often cannot guarantee the
choosing the demand influencing factors of bike sharing optimal model of building; therefore, this paper uses grid
system, we take short-term influence factors as the principal optimization method in the set interval (𝑐 ∈ [1, 5], 𝑔 ∈
thing. The preliminary influence factors are seasons, holidays, (0, 0.14)) for the parameters of support vector machine
weekends, temperature, rainfall, weather, wind, special case, (SVM) to search 𝑐 and 𝑔, the final SVM model parameters
and so on. In order to avoid the limitation of single model, for “𝑐 = 1, 𝑔 = 0.01.”
this paper chooses MLP (neural network), support vector In linear regression model, parameter selection needs to
machine, and regression model to predict the daily traffic meet the lowest possible residual and few variables, measure
volume of Zhongshan bike sharing system, respectively. The is adjusted 𝑅 (adjusted 𝑅 square), the daily usage of bike
architecture of MLP is shown below: Input signal 𝑋𝑛 through sharing system was predicted by linear regression model, and
intermediate nodes (hidden) applied to the output nodes, 𝑅 is 0.821 and reflects a good imitative effect.
through a nonlinear transformation, produce output signals Table 3 lists the main influence factors and the results of
𝑋𝑛+1 , each sample including the training of the input vector regression analysis; there are many factors that can influence
𝑋𝑛 and expected output 𝑡, network output value 𝑋𝑛+1 , and 𝑡, the effect of linear regression, but too many factors may
adjust the coupling strength between the input node (output reduce the prediction accuracy; therefore, in the experiment,
node) and hidden layer nodes, and, after repeated training, we selected the 15 influencing factors, including weather,
output the minimum error information. This paper uses temperature, and date, after comparing the influence of
the relative error as evaluation index of the whole demand different factors, and finally chose the 7 main factors as shown
prediction and chooses the optimal one from three results of in Table 3. It selected the different variables such as spring and
three methods. autumn, weekday, temperature, weather, quality, and special
The forecasting results of three methods illustrated in events. Special event includes bad weather and the Spring
Table 2. Table 2 shows the prediction result of regression Festival.
 2629, 2018, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2018/8028714 by Turkey Cochrane Evidence Aid, Wiley Online Library on [17/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 Mathematical Problems in Engineering

Table 3: The result of regression model.

Variable Coefficient 𝑡 Sig.

Constant - 16.430 .000
Spring .217 8.472 .000
Autumn .188 7.540 .000
Working day .309 12.875 .000
Temperature .067 2.761 .006
The quality of weather .212 8.229 .000
Special case −.568 −22.384 .000

Table 4: The prediction error of different types of date.

Normal day
The type of day Special day
Nonworkdays Workday
The amount of data 23 days 62 days 139 days
The average relative error of all stations
Production 53.12% 29.36% 23.69%
Attraction 50.81% 27.56% 22.15%
“Nonworkdays” include normal weekends and national statutory holidays; another date is workday.

The prediction of bike sharing system’s daily demand in 14000

2014 as shown in Figure 2 includes 224 days’ forecasting 12000
Traffic volume

results, compared with real value. The two curves of predicted 10000
8000
values and actual values fit well, reflecting a better prediction. 6000
4000
3.3. The Demand Prediction of Stations. After acquiring the 2000
forecasting results of system’s daily trip volume, we forecast 0
the production and attraction of each station. To understand 1 31 61 91 121 151 181 211
the proposed model to predict effects at different times, (Day)
the date will be divided into two categories: special date Actual value
and normal date. Special date refers to the date when a Prediction value
particular event occurs, for instance, the Spring Festival or
some extreme weather. In this case, the public bike system Figure 2: The predictable results of system’s daily demand in 2014.
is abnormal. As for the normal date, it will be divided into
a working day and a nonworking day to test the effect of
differences in demand forecasting. year, which shows the error condition of the prediction results
In this paper, we use the real data of 2013 as the original and the original results in the same period. Figure 3 shows the
input data of 𝑋(𝑡−1) ; we predict the production and attraction average relative error of daily demand prediction for stations;
of each site within 2014.1.1∼2014.8.12, which includes 23 most error value is about 20%, but some abnormal errors will
special dates and 201 normal dates (139 working days and 62 be higher than 50% and even more.
nonworking days). With the relative error as the predicting All the above is the overall analysis of traffic volume
evaluation index, we forecast the traffic volume of 167 sites in prediction. Next, this paper will analyze the prediction of a
the city of Zhongshan. single site, making readers have a more clear understanding
First, we analyze the overall predicting results. Evaluation of the prediction effect. We make an average of 224 days for
index is the average relative error for all sites, as shown in predicting error on a single site and classified statistics, as
Table 4. It shows that the average relative error of special dates shown in Table 5. The prediction error distribution of each
under abnormal circumstances is up to 50%. The average site’s production and attraction is shown in Figure 4. Table 5
relative error within the normal dates is less than 30% for directly reflects the fact that the relative error of actual volume
all sites. Prediction accuracy of working days is higher than and prediction has a negative correlation in this paper. The
that of nonworking days. This is directly associated with relative error of the site with a large traffic volume is generally
more commuting of working days. Interestingly, the average less than 20%, which has a great predicting effect. Combined
relative error of all sites’ attraction is slightly lower than the with Table 5 and Figure 4, the relative error of the site
amount of production. It may be interpreted as a user’s travel which located in city periphery with a small traffic volume is
habits. generally higher than 30%; however, the absolute error of the
As an auxiliary illustration, we use data in 2013 as the prediction of these sites is smaller, and the predicting results
original data, prediction of 2014 data, and the timespan is one actually meet the requirements.
 2629, 2018, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2018/8028714 by Turkey Cochrane Evidence Aid, Wiley Online Library on [17/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Mathematical Problems in Engineering 7

Table 5: The statistical analysis of stations in different level relative error.

Production Attraction
The average
relative error The amount of the average The amount of the average
PP PA
station absolute error station absolute error
<0.2 83 70.00% 12.40 69 60.21% 13.16
0.2–0.3 55 21.94% 8.91 77 33.20% 8.98
>0.3 29 8.07% 6.57 21 6.59% 6.44
Summation 167 100% - 167 100% -
PP: the proportion of these stations’ production in the system; PA: the proportion of these stations’ attraction in the system.

2 2

The average relative error

The average relative error

1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
1 31 61 91 121 151 181 211 1 31 61 91 121 151 181 211
(Day) (Day)
(a) Prediction error of stations’ production (b) Prediction error of stations’ attraction

Figure 3: The distribution of the average relative error of all stations at all days.

Figure 4: The distribution of predictive error of stations’ production and attraction.

At present, there is a lack of standardized evaluation Markov chain model of site demand, outperforming already
system for public bike system, which directly affects the developed solutions. With the model performance especially
judgment of the results of the forecast. A reasonable and in the study of Zhongshan bike sharing system, we find the
effective evaluation system needs more research. In this following conclusion.
paper, the relative error is the main evaluation index and First of all, the whole demand and station-level demand
the absolute error is added as a supplementary explanation, have something different in bike sharing system. For the
which can be more objective for judging the predicting whole demand forecasting, the regression model can work
results. well but cannot satisfy the station-level demand’s claim.
To ensure good predictions, the station-level demand fore-
4. Conclusions and Future Work casting needs to consider not only the most significant
influence factors, but also the constraints among stations
Effective demand forecasting is very important for bike which few papers have noticed. And this paper did it with
sharing system planning and the daily operation management good prediction result.
based on the analysis of the recent public bike demand Second, traffic flows among stations are very uneven, so
forecasting research. We built a prediction system based on the single evaluation index can hardly reflect the complete
 2629, 2018, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2018/8028714 by Turkey Cochrane Evidence Aid, Wiley Online Library on [17/05/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8 Mathematical Problems in Engineering

forecasting result, either the relative error or absolute error. [7] T. D. Tran, N. Ovtracht, and B. F. D’Arcier, “Modeling bike
Due to the differences in public bicycle travel behavior of sharing system using built environment factors,” in Proceedings
different cities, it is difficult to form a unified evaluation of the 7th CIRP Industrial Product-Service Systems Conference,
method and evaluation index. This paper based on the relative IPSS 2015, pp. 293–298, France, May 2015.
error and absolute error as the main evaluation index built [8] I. Frade and A. Ribeiro, “Bicycle Sharing Systems Demand,”
the demand forecasting evaluation method of Zhongshan Procedia - Social and Behavioral Sciences, vol. 111, pp. 518–527,
City public bike system. More data is needed to establish a 2014.
standardized evaluation process; it will be done in a future [9] Q. Chen and T. Sun, “A model for the layout of bike stations in
job. public bike-sharing systems,” Journal of Advanced Transporta-
Finally, the hybrid model represents an important devel- tion, vol. 49, no. 8, pp. 884–900, 2015.
opment direction of demand forecasting methods, which [10] P. Borgnat, P. Abry, P. Flandrin, C. Robardet, J.-B. Rouquier, and
may eliminate inherent defects of single model and withhold E. Fleury, “Shared bicycles in a city: A signal processing and data
analysis perspective,” Advances in Complex Systems (ACS), vol.
advantages of various models. In the future, we will try to
14, no. 3, pp. 415–438, 2011.
take advantage of the hybrid model to correct the forecasting
[11] H. Fanaee-T and J. Gama, “Event labeling combining ensemble
method of this paper.
detectors and background knowledge,” Progress in Artificial
Simultaneously, we will make efforts in dynamic demand Intelligence, vol. 2, no. 2-3, pp. 113–127, 2014.
forecasting of bike sharing system in next work.
[12] R. Giot and R. Cherrier, “Predicting bikeshare system usage up
to one day ahead,” in Proceedings of the 2014 IEEE Symposium
Conflicts of Interest on Computational Intelligence in Vehicles and Transportation
Systems, CIVTS 2014, pp. 22–29, USA, December 2014.
The authors declare that they have no conflicts of interest. [13] S. M. Salaken, M. A. Hosen, A. Khosravi, and S. Nahavandi,
“Forecasting Bike Sharing Demand Using Fuzzy Inference
Acknowledgments Mechanism,” in Neural Information Processing, vol. 9491 of
Lecture Notes in Computer Science, pp. 567–574, Springer
This work was supported by National Natural Science Foun- International Publishing, Cham, 2015.
dation of China under Grant no. 51178403, the Funda- [14] H. Xu, J. Ying, H. Wu, and F. Lin, “Public bicycle traffic flow
mental Research Funds for the Central Universities (no. prediction based on a hybrid model,” Applied Mathematics &
SWJTU11CX080 and no. 2682014CX130), Program for New Information Sciences, vol. 7, no. 2, pp. 667–674, 2013.
Century Excellent Talents in University (NCET-13-0977), [15] D. Singhvi, S. Singhvi, P. I. Frazier et al., “Predicting bike usage
Chengdu Science and Technology Bureau (no. 2014-RK00- for New York city’s bike sharing system,” in Proceedings of the
00034-ZF), Science & Technology Department of Sichuan 29th AAAI Conference on Artificial Intelligence, AAAI 2015, pp.
Province (no. 2014RZ0037), and the Science and Technology 110–114, usa, January 2015.
Innovation Practice Program for Graduate Student, South- [16] R. Rixey, “Station-level forecasting of bikesharing ridership,”
west Jiaotong University (no. YC201507105). Transportation Research Record, no. 2387, pp. 46–55, 2013.
[17] A. Kaltenbrunner, R. Meza, J. Grivolla, J. Codina, and R. Banchs,
References “Urban cycles and mobility patterns: Exploring and predicting
trends in a bicycle-based public transport system,” Pervasive
[1] S. Susan, G. Stacey, and Z. Hua, “Bike Sharing in Europe, the and Mobile Computing, vol. 6, no. 4, pp. 455–466, 2010.
Americas, and Asia,” Transportation Research Record, pp. 159– [18] J. W. Yoon, F. Pinelli, and F. Calabrese, “Cityride: A predictive
167, 2010. bike sharing journey advisor,” in Proceedings of the 2012 IEEE
[2] T. Yang, P. Haixiao, and S. Qing, “Bike-sharing systems in 13th International Conference on Mobile Data Management,
Beijing, Shanghai and Hangzhou and their impact on travel MDM 2012, pp. 306–311, India, July 2012.
behaviour,” in Proceedings of the Transportation Research Board [19] L. Cagliero, T. Cerquitelli, S. Chiusano, P. Garza, and X. Xiao,
Annual Meeting, 2011. “Predicting critical conditions in bicycle sharing systems,”
[3] C. Kloimüllner, P. Papazek, B. Hu, and G. R. Raidl, “Balancing Computing, vol. 99, no. 1, pp. 39–57, 2017.
bicycle sharing systems: an approach for the dynamic case,” [20] J. Froehlich, J. Neumann, and N. Oliver, “Sensing and predicting
in Evolutionary computation in combinatorial optimization, vol. the pulse of the city through shared bicycling,” in Proceedings of
8600 of Lecture Notes in Comput. Sci., pp. 73–84, Springer, the 21st International Joint Conference on Artificial Intelligence,
Heidelberg, 2014. IJCAI-09, pp. 1420–1426, usa, July 2009.
[4] X. N. Liu, J. J. Wang, and T. F. Zhang, “A method of bike sharing [21] N. Fournier, E. Christofa, and M. A. Knodler, “A sinusoidal
demand forecasting,” Applied Mechanics and Materials, vol. 587- model for seasonal bicycle demand estimation,” Transportation
589, pp. 1813–1816, 2014. Research Part D: Transport and Environment, vol. 50, pp. 154–
[5] W. El-Assi, M. Salah Mahmoud, and K. Nurul Habib, “Effects 169, 2017.
of built environment and weather on bike sharing demand: a
station level analysis of commercial bike sharing in Toronto,”
Transportation, vol. 44, no. 3, pp. 589–613, 2017.
[6] A. Faghih-Imani, N. Eluru, A. M. El-Geneidy, M. Rabbat, and
U. Haq, “How land-use and urban form impact bicycle flows:
Evidence from the bicycle-sharing system (BIXI) in Montreal,”
Journal of Transport Geography, vol. 41, pp. 306–314, 2014.