Dynamic Demand Prediction for Expanding Electric Vehicle Sharing Systems: A Graph Sequence Learning Approach

Abstract

Electric Vehicle (EV) sharing systems have recently experienced unprecedented growth across the globe. During their fast expansion, one fundamental determinant for success is the capability of dynamically predicting the demand of stations as the entire system is evolving continuously. There are several challenges in this dynamic demand prediction problem. Firstly, unlike most of the existing work which predicts demand only for static systems or at few stages of expansion, in the real world we often need to predict the demand as or even before stations are being deployed or closed, to provide information and support for decision making. Secondly, for the stations to be deployed, there is no historical record or additional mobility data available to help the prediction of their demand. Finally, the impact of deploying/closing stations to the remaining stations in the system can be very complex. To address these challenges, in this paper we propose a novel dynamic demand prediction approach based on graph sequence learning, which is able to model the dynamics during the system expansion and predict demand accordingly. We use a local temporal encoding process to handle the available historical data at individual stations, and a dynamic spatial encoding process to take correlations between stations into account with graph convolutional neural networks. The encoded features are fed to a multi-scale prediction network, which forecasts both the long-term expected demand of the stations and their instant demand in the near future. We evaluate the proposed approach on real-world data collected from a major EV sharing platform in Shanghai for one year. Experimental results demonstrate that our approach significantly outperforms the state of the art, showing up to three-fold performance gain in predicting demand for the rapidly expanding EV sharing system.

Full text

Dynamic Demand Prediction for Expanding Electric Vehicle
Sharing Systems: A Graph Sequence Learning Approach
Man Luo1, Hongkai Wen1, Yi Luo2, Bowen Du1, Konstantin Klemmer1and Hongming Zhu2
1Department of Computer Science, University of Warwick, UK
2School of Software Engineering, Tongji University, China
{m.luo.1, hongkai.wen, b.du, k.klemmer}@warwick.ac.uk, {1731530, zhu_hongming}@tongji.edu.cn
ABSTRACT
Electric Vehicle (EV) sharing systems have recently experienced
unprecedented growth across the globe. Many car sharing service
providers as well as automobile manufacturers are entering this
competition by expanding both their EV eets and renting/returning
station networks, aiming to seize a share of the market and bring
car sharing to the zero emissions level. During their fast expan-
sion, one fundamental determinant for success is the capability of
dynamically predicting the demand of stations as the entire sys-
tem is evolving continuously. There are several challenges in this
dynamic demand prediction problem. Firstly, unlike most of the
existing work which predicts demand only for static systems or at
few stages of expansion, in the real world we often need to predict
the demand as or even before stations are being deployed or closed,
to provide information and support for decision making. Secondly,
for the stations to be deployed, there is no historical record or
additional mobility data available to help the prediction of their
demand. Finally, the impact of deploying/closing stations to the
remaining stations in the system can be very complex. To address
these challenges, in this paper we propose a novel dynamic demand
prediction approach based on graph sequence learning, which is
able to model the dynamics during the system expansion and pre-
dict demand accordingly. We use a local temporal encoding process
to handle the available historical data at individual stations, and
a dynamic spatial encoding process to take correlations between
stations into account with graph convolutional neural networks.
The encoded features are fed to a multi-scale prediction network,
which forecasts both the long-term expected demand of the sta-
tions and their instant demand in the near future. We evaluate the
proposed approach on real-world data collected from a major EV
sharing platform in Shanghai for one year. Experimental results
demonstrate that our approach signicantly outperforms the state
of the art, showing up to three-fold performance gain in predicting
demand for the rapidly expanding EV sharing system.
KEYWORDS
Electric Vehicle Sharing; Dynamic Demand Prediction; Expansion
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA
©2019 Association for Computing Machinery.
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00
https://doi.org/10.1145/nnnnnnn.nnnnnnn
ACM Reference Format:
Man Luo
1
, Hongkai Wen
1
, Yi Luo
2
, Bowen Du
1
, Konstantin Klemmer
1
and Hongming Zhu
2
. 2019. Dynamic Demand Prediction for Expanding
Electric Vehicle Sharing Systems: A Graph Sequence Learning Approach.
In Proceedings of KDD’19. ACM, New York, NY, USA, Article 4, 9 pages.
https://doi.org/10.1145/nnnnnnn.nnnnnnn
1 INTRODUCTION
Car sharing services have long been recognised as an environmen-
tally friendly mobility option, reducing vehicles on the road while
cutting out unnecessary CO
2
emissions. With the recent advances
in battery technologies, a new generation of car sharing services is
going one step further, by oering full electric vehicle (EV) eets
with fast expanding infrastructures in major cities, e.g. Bluecity
1
in London, WeShare
2
in Berlin, and BlueSG
3
in Singapore. Tra-
ditional car sharing providers have also started to populate their
EV eets, e.g., ZipCar seeks to provide over 9,000 full electric ve-
hicles across London by 2025
4
. According to a recent study [
14
],
the global market of EV sharing services is poised for much faster
growth in the near future, due to the incentives and regulations put
in place by governments across the world to encourage overall EV
usages.
Despite their increased popularity, the practicality and utility
of EV sharing systems still rely heavily on the infrastructure at
renting/returning stations. In particular, for systems with the need
to rapidly expand their station networks, it is paramount to be
able to dynamically predict the accurate demand as or even before
implementing any expansion strategy. This is not only the key for
the stakeholders to make informed decisions as to where and when
to deploy new stations or close the poorly performing ones, but
also of great importance to the eective operation of currently used
stations, since understanding the potential impact of proposed ex-
pansion to their demand can provide valuable insights on a number
of vital tasks such as scheduling and rebalancing.
However, this dynamic demand prediction problem is not trivial,
especially in the context of fast expanding EV sharing systems. Most
of the existing work on demand prediction [
3
,
7
,
10
,
11
,
13
,
18
] as-
sumes the stations in the system are static, or only predicts demand
after xed expansion stages [
12
]. These assumptions often collapse
in the real world. Fig. 1(a)-(c) visualise the expansion process of a
major EV sharing platform in Shanghai during 2017. We see that in
the beginning stations are scattered within limited areas, while at
the end of 2017 the entire city has been densely covered. As shown
in Fig. 1(d), within just 12 months the total number of stations
1https://www.blue-city.co.uk
2https://www.volkswagenag.com/en/news/2018/08/VW_Brand_We_Share.html
3https://www.bluesg.com.sg
4https://www.zipcar.co.uk/electric
arXiv:1903.04051v1 [cs.AI] 10 Mar 2019

KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA M. Luo et al.
(a) Station distribution in Jan.
New stations
since Jan.
(b) Station distribution in Jul.
New stations
since July
(c) Station distribution in Dec.
0
100
200
300
400
500
600
500
1000
1500
2000
2500
3000
Jan FebMar Apr May Jun Jul Aug Sep Oct NovDec
# of stations
# of stations in operation
Closed Stations
Newly Deployed Stations
Net Increase
(d) Statistics of stations.
Figure 1: The expansion process of an EV sharing system in Shanghai during the year 2017. Images better viewed in colour.
New station
A
B
C
(a)
0
1000
2000
3000
Apr May Jun Jul Aug
# of Orders
ABCNew station
(b)
New station
E
F
G
(c)
0
2000
4000
6000
Feb Mar Apr May Jun
# of Orders
EFGNew station
(d)
Figure 2: Dierent types of impact when deploying new stations to the current station network.
in operation has doubled, as each month there are continuously
hundreds of stations being deployed. In this context, predicting
demand at those newly deployed or to be deployed stations is very
challenging, since there is no sucient historical data available as
prior knowledge.
On the other hand, dynamics introduced by the expansion pro-
cess can have very complex impact on the entire system. For exam-
ple, as shown in Fig. 2, deploying stations at various places may
have completely dierent consequences. Obviously the new sta-
tion in Fig. 2(a) ‘steals’ demand from one of its neighbours in the
following months (station A, see the changes of their order num-
bers in Fig. 2(b)), because we found the new station was deployed
at a shopping centre which may attract more users. In contrast,
deploying the new station in Fig. 2(c) has increased orders of its
neighbour stations collectively (see Fig. 2(d)). This is because the
new station was deployed at the terminal on the east side of the
airport, and many users rent vehicles for convenient short-range
connections to/from stations E, F, and G which are on the west side.
This makes accurate demand prediction for the remaining stations
also very challenging in the presence of such dynamics, due to the
non-trivial impact caused by the expansion.
To address those challenges, in this paper we propose a novel
dynamic demand prediction approach, which models the expan-
sion of EV sharing systems using graph-based sequence learning,
and is able to predict the accurate demand of stations along with
the expansion process. Specically, for each station that comes in
operation, we employ a local temporal encoding process to capture
the correlations within the historical data. The extracted features
from all stations are then compiled by a dynamic spatial encoding
process, which considers the spatial dependencies between them
as multiple time-varying graphs, and fuses the station-level fea-
tures with Graph Convolutional Neural Networks (GCN). Based
on the encoded information and future expansion plan (i.e. which
stations to be deployed/closed), our prediction network predicts
station demand at multiple scales, from the instant demand in the
immediate near future, to the long term expected demand, for both
stations to be deployed and the ones remaining. Hence, the technical
contributions of this paper are as follows:
•
To the best of our knowledge, this is the rst work that
identies and formulates the dynamic demand prediction
problem in expanding electric vehicle sharing systems.
•
We propose a novel graph sequence learning approach, which
employs temporal and spatial encoding in tandem to model
the complex dynamics of the continuous system expansion.
•
We design a new multi-scale prediction network, which is
able to forecast not only the expected demand of stations in
the long term, but also the instant future demand in subse-
quent timestamps.
•
We evaluate the proposed approach on real-world data col-
lected from a major EV sharing platform for one year. Exten-
sive experiments have shown that our approach signicantly
outperforms the state of the art, oering up to three-fold
improvement in prediction accuracy.
2 PROBLEM FORMULATION
In this section, we rst introduce some key concepts used through-
out the paper, then we formulate the problem of dynamic demand
prediction and provide an overview of the proposed framework.

KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA
2.1 Preliminaries
EV Stations:
Let
si
be a station in the Electric Vehicle (EV) sharing
system. In this paper, we assume
si
can be represented as a tuple
(xi,mi), where xiare the coordinates (e.g. latitude and longitude)
of
si
, and
mi
is the number of charging docks within
si
. We also
assume that for a given
si
, we can extract a number of geospatial
features based on its location
xi
, such as nearby Points of Interest
(POI) or the distribution of road networks within a certain radius.
Instant Station Demand:
We dene the instant demand of station
si
at timestamp
t
as the rent/return frequency of
si
when it is avail-
able, denoted as
di(t)
. In this paper the granularity of timestamp
t
is
days, i.e., we focus on daily station demand, but it is straightforward
to adopt other time granularity levels in our framework.
Expected Station Demand:
For a station
si
, the expected demand
¯
di
over a period
[ts,te]
can be dened as the mean
¯
di(ts,te)=
|te−ts|−1Íte
t=tsdi(t)
. In practice, we often consider the expected
demand from current time
t
towards the future, and aggregate it
according to some index, e.g., days of the week. Without loss of
generality, in this paper we denote the future expected demand of
station
si
as
¯
di=[¯
dMo
i,¯
dTu
i, . . ., ¯
dSu
i]
for dierent days of the week.
Station Network:
The stations of the EV sharing system can be
modelled as a graph
G=(S,A)
, where the nodes
si∈S
are stations
as dened above. An edge
ai j ∈A
may encode a certain type of
correlation between two stations
si
and
sj
, e.g., the spatial distance
between them, or similarity between their POI/road network fea-
tures. Sec. 3 will discuss how our approach constructs multiple
graphs to capture such inter-station relationships in more details.
Station Network Dynamics:
Unlike existing work, in this pa-
per we assume the station network is continuously evolving over
time. More specically, let
Gt−1=(St−1,At−1)
represents the sta-
tion network at time
t−
1. We assume at time
t−
1, there is an
expansion plan to be implemented before time
t
, which shall ex-
pand the current station network from
Gt−1
to the planned network
GP
t−1
. Let’s assume during this a set of new stations
S+
will be
deployed, while existing stations
S−
will be removed. If the ex-
pansion plan goes through, then at time
t
the station network
Gt
becomes
GP
t−1
, where
Gt=(St,At)
,
St=(St−1−S−) ∪ S+
and
At=(At−1− {aij |si∈S−or sj∈S−}) ∪ {ai j |si∈S+or sj∈S+}.
2.2 Dynamic Demand Prediction Problem
Suppose that at time
t
, we have the previous topology
G1, . . .Gt
and
demand
D1, . . ., Dt
of the station network, where
Dt={di(t)|si∈
Gt}
. Let
GP
t
be the planned station network. The dynamic demand
prediction problem tackled in this paper is that given the historical
data, for an arbitrary station in the planned network
si∈GP
t
(deployed or not yet deployed) we aim to estimate both its expected
future demand
ˆ
¯
di
and the subsequent
k
instant demand
[ˆ
di(t+
1
),ˆ
di(t+
2
), . . ., ˆ
di(t+k)]
, which minimise the mean square errors
with respect to the ground truth ¯
diand di:
δ¯
di
=|¯
di|−1∥ˆ
¯
di−¯
di∥2, and δdi=k−1
t+k
Õ
τ=t+1
∥ˆ
di(τ) − di(τ)∥2(1)
In practice, the expected demand
ˆ
¯
di
can be viewed as a metric for
the long-term performance of stations
si
, e.g., if
si
is a station to
be deployed,
ˆ
¯
di
quanties the average level of demand it may be
LSTMs at
Individual Stations
Historical
Demand
Days of
Week
Weather
Condition
Events,
Holidays
Charging
Docks #
Local POI
Local
Road Net
Pairwise
Distance
POI Distr.
Sim.
Road Net
Sim.
Multi-graph
Construction
Dynamic GCN Context
Generation
Expected
Demand
Instant
Demand
Fully Connected
Network
Decoder LSTM
with Attention
…
… …
Local Temporal Encoding Dynamic Spatial Encoding Multi-scale Demand Prediction
Local features
Spatial
correlations
Figure 3: Overview of the proposed framework for dynamic
demand prediction.
able to attract. On the other hand, the sequence of instant demand
[ˆ
di(t+
1
),ˆ
di(t+
2
), . . ., ˆ
di(t+k)]
describes the immediate trend of
station demand under the impact of the expansion plan, which can
help to optimise key future operation strategies such as marketing
and resource allocation.
2.3 Framework Overview
Fig. 3 shows the overview of the proposed framework for dynamic
demand prediction, which consists of three major components:
Local Temporal Encoding:
During the life cycle of a station
si
(from being deployed to shut down), its demand can be viewed as a
time series, where the current demand
di(t)
should correlate with
the local historical demand
di(t−
1
), . . ., di(
1
)
. In addition, there
may exist other temporal factors that can inuence the demand
of individual stations, such as weather conditions, air pollution
levels, days of the week and public holidays etc. To model such
temporal dependencies, we assign a Long Short-Term Memory
(LSTM) network at each individual station when being deployed,
and use them to encode local temporal information at station level.
Dynamic Spatial Encoding:
Intuitively, the demand of a station
si
should also be aected by the others in the station network.
To capture the spatial correlations, at each time
t
we construct
multiple graphs to encode dierent spatial relationships between
the stations, e.g., inter-station distances, POI similarity, and road
network distributions. Then we use graph convolutional neural
networks (GCN) to fuse those graphs and encode the previously
computed local features of individual stations. In particular, as
the station network is evolving over time, we consider a dynamic
version of GCN which is able to process such time-varying graphs.
Multi-scale Demand Prediction:
Based on the results of the
above temporal and spatial encoding, we aim to predict both the
expected demand and subsequent instant demand of stations after
the planned expansion. To achieve that, we design a multi-scale
prediction network, which rstly compiles the previously learned
features into a context vector. For expected demand, it uses a fully
connected branch to perform the prediction, while on the other
hand, it considers a decoder LSTM network with attention mecha-
nism to forecast instant demand at multiple future timestamps.
We are now in a position to elaborate the proposed dynamic
demand prediction approach in more detail.

KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA M. Luo et al.
Fully Connected
Decoder LSTM
Multi-graph
Dynamic GCN
Local LSTMs
t - 2 t - 1 t
Additional info.
Planned stations
Existing stations
Distance
POI
Road Net.
Distance
POI
Road Net.
Distance
POI
Road Net.
t + 1 t + 2 t + 3
Multi-scale
Prediction Networks Expected demand
Instant demand
…
…
…
…
…Planned stations
Existing stations
Planned stations
Existing stations
Figure 4: The workow of the proposed dynamic demand prediction approach.
3 METHODOLOGY
3.1 Local Temporal Encoding
Like in many other shared mobility systems, we observe that the
demand of stations in the EV sharing platform exhibits strong tem-
poral correlations, as shown later in Fig. 6(b). For instance, although
it uctuates largely over time, the demand at a station approximates
certain periodical patterns at dierent days across the week. In that
sense, exploiting such knowledge can help signicantly in estimat-
ing the accurate future demand of current stations, which will have
a positive knock-on eect when predicting demand for new stations
during expansion. However, those demand patterns are typically
inuenced by multiple complex factors such as weather, air quality
and events, and individual stations may react to those factors very
dierently. Therefore, it is often not optimal to only incorporate the
temporal information globally for the station network, but instead
in this paper we model such microdynamics at station level.
Concretely, when a station
si
is deployed, we instantiate a local
LSTM network which keeps processing its demand records and the
additional temporal information available, e.g. weather, days of the
week and public holiday/events. In our implementation, we train the
LSTMs with shared weights across stations. Then at a later time
t
,
the LSTM encodes the station’s historical demand
di(t),di(t−
1
), . . .
of
si
as well as the auxiliary information into a temporal feature
vector
fi(t)
. Moreover, in this paper we also condition
fi(t)
with a
static station feature
ci
, which describes key attributes of
si
such
as its number of available charging docks
mi
, nearby POIs and
environmental characteristics etc. Therefore,
fi(t)
and
ci
carry
important local information about individual stations since they
started operating, which are then passed on as the input for spatial
encoding. Fig. 4 shows the workow of the proposed approach,
where we see that at each timestamp we maintain a collection of
local LSTMs to encode information of individual stations.
3.2 Dynamic Spatial Encoding
3.2.1 Constructing Multiple Graphs. As discussed in Sec. 2.1, at
a given time
t
we represent the station network as a graph
Gt=
(St,At)
, where
St
are the set of current stations and
At
is the
adjacent matrix describing the pairwise correlations between them.
In practice there are often more than one types of correlations,
which can’t be eectively captured by a single graph. Therefore in
this paper we construct multiple graphs to encode the complex inter-
station relationships, particularly the distance graph, the functional
similarity graph, and the road accessibility graph (see Fig. 4).
Distance:
In most cases, we observe that the demand of stations
close to each other are highly correlated, e.g. they may be deployed
around the same shopping centre, and thus tend to be used inter-
changeably. We capture such correlations with a distance graph
AD, whose elements are the reciprocal of station distance:
aD
i j =∥xi−xj∥−1
2(2)
where
xi
,
xj
are the station coordinates, and
∥·∥2
is the Euclidean
distance. We also set diag (AD)to 1 to include self loops.
Functional Similarity:
Intuitively, stations deployed in areas with
similar functionalities should share comparable demand patterns.
For instance, stations close to university campuses typically have
signicantly higher demand during weekends. We characterise the
functionalities of stations by considering the distributions of their
surrounding POIs. Suppose we have
P
dierent categories of POIs
in total, and let
pi
be the distribution of the
P
types of POIs within
a certain radius of station
si
. The functional similarity graph
AF
is
then dened as:
aF
i j =sim (pi,pj)(3)
where
sim () ∈ [
0
,
1
]
is a similarity measure which quanties the
distance between feature vectors. In our experiments, we use the
soft cosine function.
Road Accessibility:
Another important factor that aects station
demand is the accessibility to road networks. Intuitively, stations
close to major ring roads, or within areas that have densely con-
nected streets would have higher demand. To model this, we con-
sider the drivable streets in the vicinity of a station
si
as a local
road network, containing dierent types of road segments and their
junctions. We exact a feature vector
ri
from the local road network,
which encodes information such as the road segments density, aver-
age junction degree and mean centrality etc. Given those features,
the road accessibility graph can be dened with certain similarity
function sim ():
aR
i j =sim (ri,rj)(4)
3.2.2 Dynamic Multi-graph Convolution. At time
t−
1, given the
constructed graphs
At−1={AD
t−1,AF
t−1,AR
t−1}
which describe the
inter-station relationships, we propose a dynamic multi-graph GCN

KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA
to fuse such spatial knowledge with local features
fi(t−
1
)
and
ci
computed by the station-level temporal encoding. We perform
multi-graph convolution as follows:
H(l)
t−1=σÕ
At−1∈At−1
f(At−1)H(l−1)
t−1W(l−1)
t−1(5)
where
Hl−1
t−1
and
Hl
t−1
are the hidden features in layers
l−
1and
l
respectively, while
Wl−1
t−1∈RUl−1×Ul
is the feature transformation
matrix learned through end-to-end training. In particular, the in-
put
H(0)
t−1
is the collection of local features computed at individual
stations.
f(At−1)
is a function on graphs
At−1
, e.g. the symmet-
ric normalized Laplacian [
8
] or
k
-order polynomial function of
Laplacian [
6
], and
σ
is a non-linear activation function such as
ReLU.
As discussed before, in our case the station network evolves over
time, i.e. new/existing stations can be opened/closed at any time.
For simplicity, suppose at
t
there is only one new station
s+
has
been deployed. To capture that, we recalculate the inter-station
graphs
At−1
by appending new rows and columns to them, where
the new graphs
At
contain pairwise correlations between the new
s+
and each existing stations. Note that the GCN input also changes,
i.e. now
H(0)
t
has an extra feature for this newly deployed station,
computed by the local encoding process.
On the other hand, let
sj
be the station closed at time
t
. In our
implementation, instead of removing elements from the graphs,
we simply apply a mask of zeros to the corresponding rows and
columns of
At
, and set the
j
-th row of the input
H(0)
t
to zeros
since there won’t be local features generated from
sj
anymore.
The intuition is that in our graph representation,
a:,j=
0means
station
sj
has no correlation with any other station at all, and thus
won’t propagate information in the graph convolution. In addition,
note that although
f(At)
produces lters with the same size of the
feature
H(l)
t
at each layer
l
, Eq.
(5)
can still be viewed as a local
convolution given the graphs
At
. The reason is that by denition
many elements in
At
are near zero (e.g. in the distance graph
AD
t
),
i.e. for a given station, it will be only aected by features of stations
with suciently high correlations (large non-zero elements in
At
)
with it. Conceptually, the dynamic GCN operates on snapshots of
the inter-station graphs which are constructed on-the-y, and fuses
the local temporal features at individual stations with the spatial
dependencies encoded in those graphs.
3.3 Multi-scale Demand Prediction
As discussed in Sec. 2.2, the dynamic prediction problem addressed
in this paper is to forecast the future demand of arbitrary EV sharing
stations under the planned expansion, given the historical data and
previous dynamics of the station network. We have shown in the
previous sections how we use local LSTM and GCN to encode the
spatial-temporal dynamics of the system, and in this section we
explain how to make predictions at multiple scales based on the
knowledge extracted from the models. Fig. 5 shows the architecture
of the proposed multi-scale demand prediction network.
3.3.1 Predicting Expected Demand. Let
GP
t
be the planned station
network at time
t
. Without loss of generality, we assume that com-
paring to the current network
Gt
we will deploy a candidate new
su-1 su
D(u)D(u-1)
……
Ht’Ht-1Ht-2
…
D
Cu
au,… au,t-2 au,t-1 au,t
Seq. of Instant Demand Expected Demand
Seq. of Encoded Features
Attention
Decoder LSTM
FC Network
Figure 5: The proposed multi-scale demand prediction net-
work. Left: Decoder LSTM with attention mechanism for in-
stant demand prediction. Right: Fully connected network
for expected demand prediction.
station
sN
, while we close an existing one
sj
. The goal is to predict
the future demand of stations in
GP
t
. We process this planned net-
work
GP
t
with the same approach as discussed in previous sections.
Note that there is no historical data for station
sN
since it is not
deployed yet, and therefore in local encoding we only construct
its static features
csN
, while keeping
fsN(t)
as zeros. Then we ap-
ply the same update to the inter-station graphs as discussed in
Sec. 3.2.2 (adding and masking the corresponding rows/columns),
and pass the new input
H(0)′
t
(containing features of
sN
) through
the multi-graph GCN, producing an output
H′
t
. We consider this
H′
t
as the context for prediction, since it encodes both historical
data of existing stations and information on the new candidate
station sN, together with their spatial correlations.
In this paper, we consider the expected demand of station
si
over
dierent days of the week, i.e.
¯
di=[¯
dMo
i,¯
dTu
i, . . ., ¯
dSu
i]
. To predict
¯
di
, we plug in a fully connected network to the context vector
H′
t
,
which is trained to output the future expected demand for each sta-
tion in the network
GP
t
. For the station
sN
, the predicted expected
demand of itself and nearby stations indicate the potential benets
of deploying
sN
to the current station network. In Sec. 4.3 we will
show that in real-world experiments our approach signicantly
outperforms the existing techniques in prediction accuracy.
3.3.2 Predicting Instant Demand. We also predict the future instant
demand of stations in
GP
t
over a certain time window
[t+
1
, . . ., t+k]
.
This is also of great importance in practice, especially for the station
sN
, since it forecasts the immediate impact and future trends of
the station network once
sN
is in operation. However it is more
challenging than predicting the expected demand, because essen-
tially for each station we need to predict a sequence of
k
concrete
demand instead of the aggregated values.
To address that, we design a decoder LSTM network with atten-
tion architecture, which takes the sequence of features computed
by the dynamic multi-graph GCN as input, and estimates the fu-
ture
k
instant demand. In this case, conceptually the prediction
framework becomes an encoder-decoder architecture, where the
processes of local temporal encoding and dynamic spatial encod-
ing serve together as the encoder. Let
[Ht−n, .. ., Ht−1,H′
t]
be the
sequence of features generated by our GCN. Unlike in the previ-
ous case where we only consider the last output feature
H′
t
as the
context for prediction, here for each timestamp
u
in the prediction

KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA M. Luo et al.
window
[t+
1
, . . ., t+k]
, we construct the context by fusing the
feature sequence with attention mechanism:
Cu=
t
Õ
v=t−n
αuvHv(6)
where
αuv
are the attention weights determining the contribution
of a feature
Hv
(
v∈ [t−n,t]
) in predicting the demand at time
t+u
.
Those weights
αuv
are trained through back propagation in the end-
to-end optimisation. Then the decoder LSTM consumes the context
vectors and predicts the
k
subsequent future demand. We found
in our experiments that the attention mechanism is very helpful,
since the station demand patterns tend to have strong periodic
components, e.g., demand on this Monday is highly correlated with
previous Mondays, and a single context vector is too compressed
to encode such correlation.
4 EVALUATION
In this section, we evaluate the performance of the proposed dy-
namic demand prediction approach on a real electric vehicle sharing
platform in Shanghai, China. We describe the datasets and baseline
approaches considered in our experiments (Sec. 4.1 and 4.2), and
then discuss the experimental results in Sec. 4.3.
4.1 Datasets
Electric Vehicle (EV) Sharing Data:
Our EV data is collected
from real-world operational records of an EV sharing platform
for one year (Jan. to Dec. 2017), containing information on its
renting/returning orders, and the detailed expansion process of the
station network. In particular, there were 1705 stations and 4725
electric vehicles at the beginning of 2017, while as of Dec 2017 it
had 3127 stations with a eet of 16148 vehicles in operation. In total,
the raw data contains 6,843,737 records, which were generated by
approximately 0.36 million users. Fig. 6(a) visualises the spatial
distribution of the orders (represented as lines between pick up and
return stations) in a month. Fig. 6(b) shows the number of orders in
dierent days over a month, which exhibits clear periodic patterns
with peaks in weekends.
POI Data:
We also collect the Point Of Interest (POI) date from an
online map service provider in China. In total we have extracted
4,126,844 POI entries in Shanghai, each of which consists of a GPS
coordinate and a category label. In our experiments, for each station
we consider the POIs within 1km radius. Table. 1 shows the statistics
of some POI categories.
Road Network Data:
We extract road network data in Shanghai
using OSMnx [
1
] from OpenStreetMap, which is formatted as a
graph (visualised in Fig. 6(c)). Similar with the POIs, we consider
the subgraphs within 1km radius of the stations. In our data, on
average a subgraph contains road segments of length 13.85km and
approximately 39 junctions, with a mean degree of 4.28.
Meteorology Data:
Finally, we collect the daily weather data
in Shanghai for 2017 from the publicly available sources. Each
record describes weather conditions of the day, which falls into
four dierent categories: sunny,overcast/foggy,drizzling/light snow
and heavy rain/snow. Fig. 6(d) shows the distribution of weather
conditions in Shanghai over the 12 months.
POI Type Number POI Type Number
Hospitals 4745 Banks 2988
Tourist attractions 2696 Companies 89,747
Gov. organizations 16,425 Higher education 6922
Airport services 126 Residences 51,089
Subway stations 1,729 Hotels 18,234
Bus stations 41,475 ... ...
Table 1: Statistics of some POI categories in our data.
4.2 Baselines and Metric
We evaluate two variants of the proposed dynamic demand predic-
tion approach respectively: 1)
DDP-Exp
, which predicts the future
expected demand of stations; and 2)
DDP-Seq
, which forecasts the
instant demand of stations in a subsequent time window. Both of
the two variants share the same local temporal and dynamic spatial
encoding processes, but they implement the two dierent branches
in demand prediction (as discussed in Sec. 3.3).
In particular, we compare our DDP-Exp with the following base-
lines:
KNN
, which uses a linear regressor to predict the expected demand
of existing stations. For the planned stations, it estimates their
demand with standard KNN, based on the similarity of features (e.g.
POIs) between them and the existing stations.
Random Forest (RF)
, which shares the similar idea as KNN, but
trains a random forest as the predictor.
Functional Zone (FZ)
, which implements the state of the art de-
mand prediction approach for system expansion in [
12
]. Note that
we don’t have taxi records in our data, but instead we directly feed
the ground truth check-in/out to favour this approach.
For DDP-Seq which computes the instant demand, we consider
three competing algorithms:
ARIMA + KNN
, which uses Auto-Regressive Integrated Moving
Average (ARIMA) [
16
] to forecast multi-step demand at existing
stations, and then uses KNN to estimate demand at new station
based on station features such as POIs.
LSTM + KNN
, which is similar with A-KNN, but trains LSTM
networks for temporal modelling.
Multi-graph GCN (MGCN)
, which implements a similar frame-
work as the state of the art in [
3
]. To perform fair comparison, here
we use our dynamic multi-graph GCN implementations that can
handle new/closed stations, and consider the same data sources as
in our approach.
For all approaches, we adopt the Root Mean Squared Error
(RMSE) and the Error Rate (ER) as the performance metric:
RMSE =v
u
t1
N
N
Õ
i=1
(ˆ
zi−zi)2, and ER =ÍN
i=1|ˆ
zi−zi|
ÍN
i=1zi
(7)
where
ˆ
zi
and
zi
are predicted and ground truth values respectively.
We implement the deep neural networks in the proposed ap-
proach with TensorFlow 1.10.0, and use the Adam optimiser with
learning rate of 0.001. The networks are trained on a single Titan
X GPU from scratch. For all approaches, we randomly select two
months of data for training while the subsequent month for testing,
and report the average performance.

KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA
(a)
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
0510 15 20 25 30
Day
# of Orders
(b) (c)
0.0
0.2
0.4
0.6
0.8
1.0
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Percentage
Sunny Overcast/Foggy
Drizzling/Light Snow Heavy Rain/Snow
(d)
Figure 6: Visualisation of data used in the experiments. (a) Spatial distribution of orders in one month. (b) Number of orders
in one month. (c) Road network in Shanghai. (d) Weather distribution of Shanghai in 2017.
0
1
2
3
4
5
6
7
8
Mon Tue Wed Thu Fri Sat Sun
RMSE
KNN RF FZ DDP-Exp
(a)
0.0
0.2
0.4
0.6
0.8
1.0
Mon Tue Wed Thu Fri Sat Sun
Error Rate
KNN RF FZ DDP-Exp
(b)
0
1
2
3
4
5
6
Existing New All
RMSE
KNN RF FZ DDP-Exp
(c)
0.0
0.2
0.4
0.6
0.8
1.0
Existing New All
Error Rate
KNN RF FZ DDP-Exp
(d)
Figure 7: Performance on predicting the expected demand. (a) RMSE and (b) ER of all stations across dierent days in the week.
(c) RMSE and (d) ER of existing vs. newly deployed stations vs. all stations averaged over all days of the week.
0
1
2
3
4
ARIMA
+KNN
LSTM
+KNN
MGCN DDP-Seq
w/o att.
DDP-Seq
w att.
RMSE
(a)
0.00
0.25
0.50
0.75
1.00
ARIMA
+KNN
LSTM
+KNN
MGCN DDP-Seq
w/o att.
DDP-Seq
w att.
Error Rate
(b)
Figure 8: Performance on predicting the instant demand. (a)
RMSE and (b) ER of the competing approaches.
4.3 Evaluation Results
Accuracy of Predicting Expected Demand:
The rst set of ex-
periments evaluate the overall accuracy when predicting the ex-
pected demand of stations. Fig. 7(a) and (b) show the RMSE and ER
of the proposed approach (DDP-Exp) and competing algorithms
over dierent days of the week. We see that comparing to naive
KNN, the random forest based approach (RF) can reduce the RMSE
by about 30% while ER by 20%. However, our approach (DDP-Exp)
performs signicantly better, and can achieve up to three times
improvement in both RMSE and ER. In particular, on average the
RMSE of DDP-Exp is approximate 1.961, which means when pre-
dicting the station’s expected demand, the value estimated by our
approach is only about
±
2 with respect to the ground truth. This
0
1
2
3
4
5
t+1 t+2 t+3 t+4 t+5 t+6 t+7
RMSE
ARIMA+KNN LSTM+KNN
MGCN DDP-Seq
(a)
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
t+1 t+2 t+3 t+4 t+5 t+6 t+7
Error Rate
ARIMA+KNN LSTM+KNN
MGCN DDP-Seq
(b)
Figure 9: (a) RMSE and (b) ER of the predicted instant de-
mand for dierent prediction lengths.
conrms that the proposed approach can eectively model the com-
plex temporal and spatial dependencies within the evolving station
network, and exploits that to make more accurate predictions. In
addition, we observe that the RMSE tends to increase on weekends
compared to weekdays for all algorithms. This is because in practice
the absolute demand on weekends is larger, which often leads to
bigger RMSE. Note that the ER remains relatively consistent across
dierent days.
Planned vs. Existing Stations:
This experiment investigates the
prediction performance of dierent approaches on the planned
stations which haven’t been deployed yet, and existing stations
which have already been in operation. Fig. 7(c) and (d) show the
average RMSE and ER of the proposed approach (DDP-Exp) and

KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA M. Luo et al.
2.0
2.5
3.0
3.5
4.0
0.0 0.2 0.4 0.6 0.8 1.0
RMSE
Exp-Demand Inst-Demand
(a)
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0
Error Rate
Exp-Demand Inst-Demand
(b)
Figure 10: Sensitivity of our approach with dierent levels
of augmented station network dynamics.
the competing algorithms on the planned, existing, and all stations
respectively. We see that all of approaches perform better on the
existing stations than the planned. This is expected because for
existing stations we have access to their historical demand data,
which is not available for planned stations. We also observe that
although the functional zone based approach (FZ) performs better
than the baselines for the planned stations, it fails on the existing
stations (performs worse than RF). This is because by design FZ is
tuned to predict demand of new stations in the context of system
expansion, but not for existing ones. Finally, we see that for both
planned and existing stations our approach (DDP-Exp) performs
consistently the best. For the planned stations, it halves the errors
comparing to the state of the art approach FZ, while for the existing
stations, it oers about three-fold improvement over the baselines.
Accuracy of Predicting Instant Demand:
This set of experi-
ments evaluates the performance of dierent approaches when
predicting the future instant demand. Here we only consider the
planned stations, since it is straightforward to predict for the ex-
isting stations given their historical data. We ask all approaches to
predict the instant demand over the next seven days, and report the
average accuracy. Fig. 8 shows the RMSE and ER of the proposed
approach (DDP-Seq) and the competing algorithms. We see that
in this very challenging case, our approach (DDP-Seq) can still
achieve an average RMSE of 2.903, which is over 30% lower than
the baselines (similar gap can be observed in ER). It is also superior
to the state of the art MGCN approach which also uses multi-graph
GCN, with about 20% reduction in RMSE and ER. This conrms that
even for the planned stations without historical data, our approach
can still accurately predict their future instant demand within a
certain time window. In addition, we nd that the attention mech-
anism in our approach is very eective. Without using attention
architecture in the decoder, the performance of our approach drops
by approximately 15%, which is still better than the state of the art.
Accuracy vs. Prediction Length:
This experiment studies the
accuracy of competing approaches when predicting instant demand
over dierent time intervals. As in the previous experiment, here we
also only consider prediction performance for the planned station.
We vary the length of the prediction time window from 1 to 7, i.e.
from predicting the demand of stations on the immediate next day
t+
1, to that on the subsequent seven days
t+
7. Fig. 9 shows the
RMSE and ER of the approaches under dierent time windows. We
observe that in general, the RMSE increases as the length of the
time window grows, especially for our approach (DDP-Seq) and the
state of the art MGCN. This makes sense because clearly predicting
demand over a longer time window is more dicult. On the other
hand, we see that the ER of baselines are higher for short window
lengths comparing to the MGCN or our approach. We nd that
this is because the baselines tend to report random estimations
on the future demand, where for shorter windows this can lead to
larger ER, but will be averaged out for longer time windows as the
ground truth demand grows in later days. Finally, we see that MGCN
can oer comparable performance with our approach (DDP-Seq)
when predicting for the immediate next timestamp. However as the
prediction length increases, our approach consistently outperforms
MGCN, with a performance gap of up to 26%.
Augmentation of Station Network Dynamics:
The last set of
experiments investigates the impact of augmenting station network
dynamics during training. As discussed in previous sections, one of
the key challenges addressed in this paper is to forecast the demand
of planned stations which haven’t been deployed, in the presence of
a continuously evolving station network. This means that we have
to explicitly learn the particular dynamics caused by deploying new
stations in order to make accuracy predictions. To account for that,
in addition to the actual dynamics within the data, during training
we articially inject dierent levels of augmented dynamics to
the station network, by simulating the process of deploying new
stations. More concretely, at each timestamp we randomly pick a
subset of existing stations according to a probability
p
, and ignore
their previous demand records, i.e. we assume that those stations
have just been deployed. We vary
p
from 0 to 1, indicating the
least (no) augmentation to the most. As shown in Fig. 10, we see
that as
p
increases from zero, our approach tends to make more
accurate predictions for both expected and instant demand. This
conrms that by injecting the augmented dynamics, we essentially
force the GCN to learn how to better react to the deployment of
new stations. We also observed that for larger
p
values, the errors
(both RMSE and ER) increase for both types of demand. This is also
expected because in those cases the excessive injected dynamics
would mute the useful information coming from local LSTMs at
individual stations and confuse the GCN, leading to deterioration
of performance. Therefore empirically we set
p
to values around
0.4∼0.6 to achieve the desired balance.
5 RELATED WORK
Demand Prediction for Shared Mobility:
Predicting user de-
mand in shared mobility services (e.g. taxi and bike- or vehicle-
sharing systems) has received considerable interest in various re-
search communities. Most of the existing work takes the histori-
cal usage (e.g. picking-up and returning records), geospatial data
such as POIs, and other auxiliary information (e.g. weather) into
account, and builds prediction models that can forecast demand
over certain periods or aggregated time slots. They also predict
the demand at dierent spatial granularity, e.g. over the entire
systems [
15
,
20
], grids/regions [
6
], station clusters [
5
,
10
,
13
], or
individual stations [
3
,
7
,
11
,
18
,
21
]. This paper falls into the last
category since we aim to predict station-level demand of EV sharing
platforms. However, our work is fundamentally dierent in that we
assume the station network is not static, but dynamically evolving,

KDD’19, Auguest 3 - 7, 2019, Anchorage, Alaska, USA
i.e. stations can be deployed or closed at arbitrary times. In this
case, state of the art station-level demand predictors (e.g. [
7
]) will
fail because they rely heavily on station historical data to make
predictions, which are not available for newly deployed stations.
Shared Mobility Expansion:
There is also a solid body of work
focusing on modeling the expansion process of shared mobility sys-
tems, e.g. planning for optimal new stations [
11
,
17
], or increasing
the capacity of existing stations [
4
]. However, all of them assume
that demand of the stations (renting and returning) are known, or
can be estimated from other data sources such as taxi records, which
is dierent from our work. On the other hand, the work in [
12
]
proposes a functional zone based hierarchical demand predictor
for shared bike systems, which can estimate the average demand
at newly deployed stations across dierent expansion stages. Our
work shares similar assumptions with [
12
], yet diers substantially:
1) instead of xed stages, we can predict demand while the entire
station network is dynamically expanding; 2) we are able to es-
timate both the instant and expected demand of new or existing
stations, while [
12
] can only predict aggregated demand patterns;
and nally 3) we don’t require historical mobility data in the newly
expanded areas, like the taxi trip records used in [12].
Graph-based Deep Learning:
Due to their non-Euclidean nature,
many real-world problems such as demand/trac/air quality fore-
casting that require spatio-temporal analysis have been tackled with
the emerging graph-based deep learning techniques [
3
,
6
,
9
,
19
]. In
particular, existing work often employs the graph convolutional
neural network [
2
] to capture the spatial correlations, where tem-
poral dependencies are typically modelled with recurrent neural
networks. For instance, [
9
] models the trac ow as a diusion
process on directed graphs for trac forecasting, while [
19
] and [
6
]
propose frameworks that use multi-graph convolutional neural
networks (CNNs) to predict demand for taxi and ride-hailing ser-
vices. Another work in [
3
] uses an encoder-decoder structure on
top of multi-graph CNNs to estimate ow between stations in bike
sharing systems, which bears a close resemblance to this paper.
However, unlike [
3
] who only output demand at the immediate
next timestamp, our work considers a sequence to sequence model
with attention mechanism to perform multi-step forecasting to-
wards future demand. In addition, none of the above approaches
can work on new stations where historical data is not available.
6 CONCLUSION
In this paper, we propose a novel dynamic demand prediction ap-
proach for expanding electric vehicle (EV) sharing systems, which
learns the complex system dynamics from the continuous expansion
process, and is able to robustly predict demand for both existing
stations and the planned stations which haven’t been deployed.
Specically, we rst encode the local temporal information at in-
dividual station level, and then fuse the extracted features with
dynamic graph convolutional neural networks (GCN) to account for
the spatial dependencies between stations. The demand of stations
is estimated by a multi-scale prediction network, which forecasts
both the long-term expected demand and the instant future demand
of the system. We evaluate our approach on data collected from a
real-world EV sharing platform for a year. Extensive experiments
have shown that our approach consistently outperforms the state
of the art in predicting both long-term expected and immediate
future demand of the fast expanding system.
REFERENCES
[1]
Geo Boeing. 2017. OSMnx: New methods for acquiring, constructing, analyzing,
and visualizing complex street networks. Computers, Environment and Urban
Systems 65 (2017), 126 – 139. https://doi.org/10.1016/j.compenvurbsys.2017.05.
004
[2]
Joan Bruna, WojciechZaremba, Arthur Szlam, and Yann LeCun. 2013. Spectral net-
works and locally connected networks on graphs. arXiv preprint arXiv:1312.6203
(2013).
[3]
Di Chai, Leye Wang, and Qiang Yang. 2018. Bike ow prediction with multi-graph
convolutional networks. In Proceedings of the 26th ACM SIGSPATIAL International
Conference on Advances in Geographic Information Systems. ACM, 397–400.
[4]
Bowen Du, Yongxin Tong, Zimu Zhou, Qian Tao, and Wenjun Zhou. 2018.
Demand-Aware Charger Planning for Electric Vehicle Sharing. In Proceedings of
the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data
Mining. ACM, 1330–1338.
[5]
Jon Froehlich, Joachim Neumann, Nuria Oliver, et al
.
2009. Sensing and predicting
the pulse of the city through shared bicycling.. In IJCAI, Vol. 9. 1420–1426.
[6]
Xu Geng, Yaguang Li, Leye Wang, Lingyu Zhang, Qiang Yang, Jieping Ye, and Yan
Liu. 2019. Spatiotemporal Multi-Graph Convolution Network for Ride-hailing
Demand Forecasting. In 2019 AAAI Conference on Articial Intelligence (AAAI’19).
[7]
Pierre Hulot, Daniel Aloise, and Sanjay Dominik Jena. 2018. Towards Station-
Level Demand Prediction for Eective Rebalancing in Bike-Sharing Systems.
In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining. ACM, 378–386.
[8]
Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classication with
Graph Convolutional Networks. In International Conference on Learning Repre-
sentations (ICLR).
[9]
Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diusion Convolutional
Recurrent Neural Network: Data-Driven Trac Forecasting. In International
Conference on Learning Representations (ICLR’18).
[10]
YexinLi, Yu Zheng, Huichu Zhang, and Lei Chen. 2015. Trac prediction in a bike-
sharing system. In Proceedings of the 23rd SIGSPATIAL International Conference
on Advances in Geographic Information Systems. ACM, 33.
[11]
Junming Liu, Qiao Li, Meng Qu, Weiwei Chen, Jingyuan Yang, Hui Xiong, Hao
Zhong, and Yanjie Fu. 2015. Station site optimization in bike sharing systems. In
Data Mining (ICDM), 2015 IEEE International Conference on. IEEE, 883–888.
[12]
Junming Liu, Leilei Sun, Qiao Li, Jingci Ming, Yanchi Liu, and Hui Xiong. 2017.
Functional zone based hierarchical demand prediction for bike system expansion.
In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining. ACM, 957–966.
[13]
Eoin O’Mahony and David B Shmoys. 2015. Data Analysis and Optimization for
(Citi) Bike Sharing.. In AAAI. 687–694.
[14]
Susan Shaheen, Adam Cohen, and Mark Jaee. 2018. Innovative Mobility: Car-
sharing Outlook. (2018).
[15]
Wen Wang. 2016. Forecasting Bike Rental Demand Using New York Citi Bike
Data. (2016).
[16]
Billy M Williams and Lester A Hoel. 2003. Modeling and forecasting vehicular
trac ow as a seasonal ARIMA process: Theoretical basis and empirical results.
Journal of transportation engineering 129, 6 (2003), 664–672.
[17]
Yanhai Xiong, Jiarui Gan, Bo An, Chunyan Miao, and Ana LC Bazzan. 2015.
Optimal Electric Vehicle Charging Station Placement.. In IJCAI. 2662–2668.
[18]
Zidong Yang, Ji Hu, Yuanchao Shu, Peng Cheng, Jiming Chen, and Thomas
Moscibroda. 2016. Mobility modeling and prediction in bike-sharing systems.
In Proceedings of the 14th Annual International Conference on Mobile Systems,
Applications, and Services. ACM, 165–178.
[19]
Huaxiu Yao, Fei Wu, Jintao Ke, Xianfeng Tang, Yitian Jia, Siyu Lu, Pinghua Gong,
Jieping Ye, and Zhenhui Li. 2018. Deep Multi-View Spatial-Temporal Network
for Taxi Demand Prediction. In 2018 AAAI Conference on Articial Intelligence
(AAAI’18).
[20]
Yu-Chun Yin, Chi-Shuen Lee, and Yu-Po Wong. 2014. Demand Prediction of
Bicycle Sharing Systems. (2014).
[21]
Ming Zeng, Tong Yu, Xiao Wang, Vincent Su, Le T Nguyen, and Ole J Mengshoel.
2016. Improving Demand Prediction in Bike Sharing System by Learning Global
Features. Machine Learning for Large Scale Transportation Systems (LSTS)@
KDD-16 (2016).