Electrical Power and Energy Systems 141 (2022) 108218

Contents lists available at ScienceDirect

International Journal of Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Data-driven intelligent EV charging operating with limited chargers

considering the charging demand forecasting
Jiayan Liu a, *, Gang Lin a, *, Christian Rehtanz a, Sunhua Huang b, Yang Zhou a, Yong Li c
a
Institute of Energy Systems, Energy Efficiency and Energy Economics, TU Dortmund, Dortmund 44227, Germany
b
School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China
c
College of Electrical and Information Engineering, Hunan University, Changsha 410082, China

A R T I C L E I N F O A B S T R A C T

Keywords: Coordinated charging scheduling can improve the operating economics of charging stations and reduce the
Electric Vehicle required amount of charging facilities. However, existing optimal scheduling schemes either simplify the
Charging Scheduling charging station capacity modeling when taking into account the traffic uncertainty, or ignore the future
Data-Driven
charging demands when considering charging capacity limitation. To tackle this issue, a data-driven intelligent
Charging Demand
EV charging scheduling algorithm is proposed in this paper, by scheduling in response to the time-of-use (TOU)
electricity price, the limitation of charging facilities, and detailed charger operating process is also considered.
First, based on the neural network algorithm, a charging demand forecasting method is introduced to establish
the charging task of the charging station. Then, according to the established task, an optimization model that
considers the charging costs, battery degradation, and users’ dissatisfaction comprehensively is proposed. The
proposed model is formulated as a mixed-integer nonlinear programming problem, and a corresponding
approach for solving the model is also proposed. Finally, the real-time operation process of the proposed
scheduling method in the actual charging station is presented. By comparing with the existing methods, better
effectiveness and performance of the proposed scheduling method are verified by simulation results.

incorporating the coordinated charging of EVs at the planning stage to

1. Introduction
With the increasing demand for energy conservation and emission
reduction, replacing fuel vehicles with electric vehicles (EVs) is a
mainstream solution to cut down the massive greenhouse gas emissions
from the transportation sector [1]. Due to the limited cruising range and
long charging time, large-scale promotion of EVs requires sufficient
charging facilities [2]. In recent years, parking lots and fast-charging
stations have attracted heavy investment, but improper coordinated
management and under-utilization of chargers increase the investment
waste and bring disappointing profitability of charging stations. In
addition, due to the larger charging power, the scale of the constructed
charging station is restricted by the power grid structure [3,4].
Recent reports [5,6] pointed out that the chargers in the charging
station have long idle with low utilization rate, (e.g. average utilization
rate of chargers in China in 2019 is less than 5%), which greatly
dampens the enthusiasm for investment in the expansion of charging
facilities. Fewer charging resources can even serve more EVs if appro­
priate charging power scheduling approaches are adopted. Therefore,
enhance the utilization rates of charging resources attracted attention.
For example, the charging spot models to share one charger to several
EVs are proposed in [7,8], and the corresponding charging station
models such as Single-to-Multiple [7] and Multiple-to-Multiple [9] are
proposed to deal with the planning problems mentioned above. The
application case [10] is also built in Shenzhen, China. The research on
charging scheduling considering charging uncertainty and limited
chargers is carried out based on this type of charging station in this
paper.
Scheduling charging power is a necessary aspect of charging station
operation that can affect charging resources utilization and users’
satisfaction positively. In order to coordinate the charging behavior of
EVs and improve energy efficiency, extensive works focusing on
charging scheduling schemes have been carried out. Since the charging
demand of EVs has obvious sequential characteristics, the optimal
scheduling method that considers the time-of-use (TOU) electricity price
has been widely adopted to promote the total utility for the charging
operator markets [10] and minimize the total charging cost and the
energy cost from the substation [11]. In addition, a scheduling method

* Corresponding authors.
E-mail addresses: jiayan.liu@tu-dortmund.de (J. Liu), Gang.lin@tu-dortmund.de (G. Lin).

https://doi.org/10.1016/j.ijepes.2022.108218
Received 3 December 2021; Received in revised form 16 February 2022; Accepted 4 April 2022
Available online 22 April 2022
0142-0615/© 2022 Elsevier Ltd. All rights reserved.
J. Liu et al.
Nomenclature
Indices
m Index of charger
n Index of EV
t Index of time slot
Δt One time slot
τ Index of current time slot
M Number of charger
Parameters
Nτ Number of EVs in VECτ
Nre τ/Nes τ Number of real/estimated EVs in VECre τ/VECim τ
Na t Number of arrived EVs at time slot t
Na,re t/ Na,es t Number of real/estimated arrived EVs at time slot t
Na,pred t Predicted number of arrived EVs at future time slot (t ∈ [τ
+ Δt,Tτ])
Nd t Number of departure EVs at time slot t
Nd,re t/ Nd,es t Number of real/estimated departure EVs at time slot
t Sreq n
Nd,pred t Predicted number of departed EVs at future time slot (t ∈
[τ + Δt,Tτ])
Ns t Number of EVs staying in the charging station at time slot t
Ns,re t/ Ns,es t Number of real/estimated EVs staying in the charging
station at time slot t
Ns,pred t Predicted number of stayed EVs at future time slot (t ∈ [τ +
Δt,Tτ])
ψt Expected total energy demand of the charging station at
time slot t
considering multiple charger types is proposed in [1]. By responding to
TOU electricity prices, users can flexibly choose the type of charger, and
the operating efficiency of charging stations is improved. For the grid
operators, TOU electricity prices reflect the state of the grid, flexible EV
charging scheduling methods can smooth the peak load and support the
valley load, which can reduce the impact of large-scale EVs charging on
the power grid [12]. In addition, scheduling the charging power tra­
jectory to the daily renewable generation curve can provide an eco­
nomic operation of the power grids [13].
From the perspective of EV users, the completion of charging and the
battery degradation rate are their most important indicators affecting
satisfaction. EV charging characteristics will impact the battery capacity
fading, which has significant effects on the battery lifetime. To investi­
gate the degradation of lithium-ion batteries in vehicle-to-grid (V2G)
programs, [14] proposed a practical wear cost model for EVs charge
scheduling applications. In [15], a comprehensive scheduling model
that takes both battery degradation and the interests of the parking
garage into consideration is devised. However, due to the limitation of
charging facilities, not all EVs can complete their charging requirements
on time. In order to satisfy charging requirements, battery degradation
in the electric vehicle charging optimization model is developed in [16].
But the charging station capacity has not been discussed yet. In [17], a
so-called on-arrival commitment scheduling strategy considering the
charging station capacity is proposed. This approach can notify each EV
whether or not it can receive the charging demand. But the limitation on
specific chargers has been simplified to the capacity of the charging
station, which means they did not consider how to specifically dispatch
the power of different chargers to each individual EV. The online pricing
mechanism for EV charging is another idea for dealing with charging
facility limitations [18,19]. By appropriately deciding charging prices,
charging operators encourage EVs to charge during free time, and reject
some contracts during peak hours. However, this method does not fully
mobilize the flexibility of charging and will force users to passively
2
International Journal of Electrical Power and Energy Systems 141 (2022) 108218
ψ re t/ψ es t Expected energy demand from the real/estimated EVs at
time slot t
ψ pred t Predicted total energy demand at future time slot (t ∈ [τ +
Δt,Tτ])
Tτ Latest departure time of the currently known EVs
Γn Set of charging information of EV n
Γ re n,τ/Γ es n,τ Set of real/estimated charging information of n-th
EV at time slot τ
VECτ Set of overall charging task at current time slot τ
VECre τ Set of real charging task at current time slot τ
VECes τ Set of estimated charging task at current time slot τ
VECre,a τ Set of EVs charging task just arriving at the station at
current time slot
VECre,un τ-Δt Set of uncompleted charging tasks at last time slot τ −
Δt
et TOU electricity price at time slot t
ta n Arrival time of EV n
td n Departure time of EV n
Sini n Initial battery SOC of EV n
Requested battery SOC of EV n
Ereq n Requested power of EV n
Eav n Average energy demand of EV n at one time slot, which is
evenly distributed according to its dwell time
Pmax Maximum power per time slot provided by one charger
Decision variables
Pn,t Decision variable, charged power of EV n at time slot t
ζn,t Binary decision variable, which is equal to 1 if EV n
charged at the time slot t, otherwise, zero
change their charging behavior, which is negative for satisfaction.
In addition to the issues discussed above, the uncertainty of future
charging demand is also considered in this paper. In the time dimension,
charging scheduling is to allocate the corresponding power to the EV
staying in the charging station at different times. From the perspective of
planning, the construction method of charging stations considering the
uncertainty of traffic flow is discussed in [20]. However, in the actual
scheduling, the uncertainty of subsequent charging demand will affect
the current scheduling performance. Therefore, the prediction of sub­
sequent charging demand is necessary to improve the efficiency of
scheduling. However, for most of the existing works, the future charging
demand is simply divided into two different scenarios to keep the
problem computationally tractable [21,22]. A charge scheduling model
considering multiple possible stages in the future is proposed in [23].
Nevertheless, it takes several hours to obtain a high-quality solution,
which is inefficient in practical applications. In [24], the charging
scheduling approach of EVs with local renewable energy under uncer­
tain EV arrival is proposed. By in responding to the grid power price, the
charging economy can be improved. But the operating process neglects
the flexibility of the scheduling charging power and battery degradation.
Artificial intelligence algorithms have broad application prospects as
tools for processing fuzzy stochastic data. A neural network based
approach for forecasting travel behavior to improve the scheduling ef­
ficiency is proposed in [25]. Based on deep reinforcement learning, [26]
proposed a real-time EV charging scheduling without establishing the
detail optimization model. But the charging demand constraints of EVs,
which can make sure the EVs can be fully charged upon departure, have
been ignored. Therefore, a constrained Markov Decision Process EV
charging scheduling approach based on safe deep reinforcement
learning is proposed by [27]. However, the methods mentioned above
all use the centralized forecasting methods, which regard the power
demand of the charged EV as a whole, ignoring the traffic flow infor­
mation of each individual EV. Therefore, those prediction and
J. Liu et al.
scheduling methods cannot be applied to the charging stations with
limited chargers that can be flexibly scheduled. The impact of a limited
number of charging facilities on scheduling is discussed in our previous
work [28]. However, to apply this scheduling method, the future
charging demand of each EV is required, which means it is only suitable
for the parking lot with a fixed charging timetable, such as electric bus
charging and electric logistics truck charging because the demand un­
certainty is ignored. Besides that, the battery degradation characteristics
that affect user satisfaction are also not considered.
Considering that the scale of charging facilities is limited by invest­
ment and grid structure, the charger utilization rate is not high. The
scheduling method in the multiple-to-multiple charging station [9] that
can improve charging flexibility is discussed in this paper. To the best of
our knowledge, there are no published works that implement the
charging scheduling considering both the charging demand uncertainty
and the operating of limited charging facilities in such form of charging
stations. The charger limit mentioned in this paper is not a simple setting
of the overall capacity limit of the charging station, but specific to which
charger serving for which EV at each time slot, and to the output
charging power of each charger at each time slot. To improve the
operating profit, it is essential for the charging operator to take both
electricity price and users satisfaction into consideration. Charging de­
mand uncertainty is also an important factor affecting the optimized
operation of the charging station, which is also discussed in this paper.
Therefore, a data-driven intelligent EV charging scheduling (DICS)
method is proposed to guarantee different EV users charging re­
quirements and improve the charging station profit, which is more
versatile than [28].
The main contributions of this paper are summarized: 1) A detailed
EV charging demand forecasting method based on neural networks is
proposed. The forecasting method can give detailed prediction results to
guide the optimization of scheduling with a limited number of chargers,
and use traffic flow data to update the prediction results in real-time. 2)
The charging scheduling optimization model considering the limited
charging facilities is proposed. Not like the other exiting works that
simplified the charger limitation into the power capacity, the proposed
scheduling method can arrange each charger to serve specific EVs with
corresponding charging power. 3) Combining the charging demand
forecasting method and the scheduling optimization model, the real-
time charging scheduling system operation process is introduced. By
considering the real and estimated EVs in the optimization model, the
more accurate guidance for the charger power allocation at the current
Fig. 1. Model of the data-based intelligent charging scheduling system.
3
International Journal of Electrical Power and Energy Systems 141 (2022) 108218
moment can be obtained. The total charging cost for the charging op­
erators is reduced while the charging requirements and reducing the
battery degradation are assured. 4) Different operation methods are
designed and compared with the proposed DICS method to verify its
effectiveness.
The rest of the paper is organized as follows: Section 2 introduces the
overview system model and the based charging cost and battery model.
Section 3 describes the EV charging demand forecasting method. The
charging scheduling model and corresponding real-time implementa­
tion method are introduced in Section 4. Section 5 shows the perfor­
mance and efficiency of the proposed DICS method and compares it with
other existing methods. The conclusion is summarized in Section 5.
2. System model
As discussed above, coordinating charging stations for EVs can help
improve the overall charging economy and reduce the construction cost
of charging stations. Therefore, the multiple-to-multiple charging sta­
tion type designed in [9] is applied for charging scheduling designing.
The proposed DICS method in this paper is mainly for operating a single
charging station and can be applied to places suitable for optimized
charging scheduling, such as work locations, commercial centers, and
residential areas. The data-based intelligent charging scheduling system
model is shown in Fig. 1. When an EV intends to charge at the charging
station, the system will report its required charging information,
including arrival time, departure time, and energy requirement. Since
the performance of the power scheduling is time-dependent and the
information of the subsequent EVs is unknown, we apply the informa­
tion of historical arrived EVs to predict the future charging demand.
With the reported and predicted EVs’ information, the intelligent
charging scheduling procedure is controlled by the charging scheduling
system (CSS). Each EV parks in the charging area and connects to the
charging network. The charging network is the multiple-charger mul­
tiple-port charging system, each charger is allowed to serve multiple
parking spaces and each EV is allowed to be charged by multiple char­
gers, but one charger can only charge one EV at a time [9]. By consid­
ering the TOU price, battery degradation characteristics, and users’
satisfaction, the CSS coordinates the service objects of each charger and
adjusts the power dispatch from each charger. All the charging activities
are automatically switched.
Fig. 2 gives a toy example about the advantage of scheduling
considering the EV information forecasting in the case of limited
J. Liu et al.
Fig. 2. Toy example for the scheduling (a) without EV information forecasting at time slot 1 (b) without EV information forecasting at time slot 3 (c) with EV
information forecasting at time slot 1 (d) with EV information forecasting at time slot 3.
charging facilities. Assume that the maximum power that a charger can
provide in a time slot is Pmax = 4 kW, and only one charger is available in
the charging station that can only serve one EV at the same time. In
Fig. 2 (a), the current time slot is 1, the EV Γ 1 with 10kWh energy de­
mand arrives at time slot 1 and will leave at time slot 5. When no sub­
sequent EVs are considered for scheduling the optimal charging scheme
is to charge 2kWh, 4kWh, 4kWh at time slots 1, 4, 5 for Γ 1. Because this
charging scheme can ensure user Γ 1 completes its charging demand
before leaving, and the TOU price at time slots 1, 4, 5 are lower than
other times. From time slots 1 to 2, the scheduling process in total
provides 2kWh energy to Γ 1 to avoid the high TOU price. However,
when the time slot comes to 3, an actual EV Γ 2 with 6 units of energy
demand arrived and will leave at time slot 5, shown in Fig. 2 (b). From
time slots 3 to 5, it can only provide a maximum of 12kWh units of
energy, while Γ 1 and Γ 2 have a total of 14kWh energy demand at this
time, so the charging demand cannot be fully guaranteed. In addition,
the charging station has to provide charging power at time slot 3 when
4
International Journal of Electrical Power and Energy Systems 141 (2022) 108218
the TOU price is the highest, which further reduces the charging
economy.
In the other case, illustrated in Fig. 2 (c), assume at time slot 1, it is
predicted that EV Γ 2 with 7kWh energy demand will arrive at time slot 4
and leave at time slot 5. Then the optimal charging scheme is to charge
Γ 1 4kWh, 4kWh, 1kWh, and 1kWh energy at time slots 1, 2, 3, and 4, the
predicated Γ 2 is charged 3kWh and 4kWh energy at time slots 4 and 5
respectively. After one charger completes the charging task of one EV, it
can start to provide charging service for another EV. In Fig. 2 (c), the Γ 1
only needs 1 kWh of battery capacity at the time slot 4. Therefore, it
needs to charge with the maximum charging power (4 kW) within only
1/4 unit time to complete the charging. Then, the remaining 3/4 unit
time of the time slot can be used to charge Γ 2. By charge with the
maximum charging power 4 kW, 3kwh energy can be charged into Γ 2 in
the remaining 3/4 unit time. Similarly, when the time slot comes to 3,
the actual EV Γ 2 with 6kWh units of energy demand arrived, shown in
Fig. 2 (d). Since the charging margin has been taken into account by
J. Liu et al. International Journal of Electrical Power and Energy Systems 141 (2022) 108218

forecasting, 8kWh energy had been charged to Γ 1 in advance at times 1 the cell voltage; D is the number of composed identical cells; βi, i = 1,
and 2. Therefore, after the actual Γ 2 arrived, both Γ 1 and Γ 2 can get …,7 are the parameters specified in [32].
sufficient energy supply at the lower TOU price period, time slots 4 and
5. 3. Data-Driven EV charging demand forecasting
In the case of limited charging facilities, we not only need to predict
the amount of energy demand in the subsequent moments but also need In the actual scene, the charging demand information of EVs that will
to predict the number of EVs and allocate the energy demand to them. arrive in the future is unknown. However, the information of the sub­
Even if the forecasting process cannot provide the completely accurate sequent EVs will affect the scheduling efficiency. Thus, it is essential to
follow-up EV data, the scheduling with subsequent can still provide a forecast the subsequent charging demand for improving scheduling ef­
certain margin for the charger and bring a significant improvement in ficiency, reducing charging costs, and increasing users’ satisfaction.
the economy and user satisfaction. Of course, the more accurate the Data Preprocessing: In order to improve the performance of intel­
forecast is, the more efficiency would be improved. ligent charging scheduling, detailed EV information (including arrival
time, departure time, and required energy) is required to be predicted.
From the perspective of the charging station operator, it is easy to record
2.1. Electric vehicle charging demand information
the number of vehicles arriving and departing at different times, as well
as the energy demand data of the respective EVs. Define Na t and Nd t are
In the realistic situation, the charging demand in a charging station is
the number of EVs arriving and departing at time slot t. Ns t denotes the
time-depended, which means at different times, the number of EVs in
number of EVs dwelling at the charging station at time slot t. Let ψ t
the charging station is different. The arrival time and departure time of
denotes the expected total energy demand at time slot t, which is
the n-th EV are ta n and td n, where ta n < td n. Ereq n represents the
computed as:
required energy of n-th EV. The required charging energy is determined
⎧
by the initial battery state of charge (SOC) Sini n and the requested SOC ⎪
⎪ Ereq
Sreq n where both of them are following the Gaussian distribution [29]. ⎪
⎪ Enav = Δt ( d n a )
⎨
tn − tn
Consequently, the required energy for each EV is computed by: (6)
⎪
⎪ ∑N
( ) ⎪
⎪ av a d
Enreq = Snreq − Snini Capn (1) ⎩ ψt = E , where ∀tn ⩽t⩽tn
n=1 n

where Capn is the battery capacity of n-th EV. Pmax is the maximum where Δt represents one time slot; Eav n means the average energy
power that a charger can provide within one time slot. It is assumed that demand of EV n at one time slot, which is evenly distributed according to
the charging demand of all cars will not exceed the maximum power that its dwell time. The ψ t represents the sum of the Eav n of all EVs staying in
can be obtained during the dwell time, which means td n − ta n ≥ Ereq n/ the station at time slot t (The EV n must stay in the charging station at
Pmax. The arrival time, departure time, and required energy of each EV time slot t, that is t ∈ [ta n, td n]).
are required information for optimal scheduling, which can be regarded Recurrent neural network layer: Recurrent neural networks
as a User, defined as: (RNN) have a powerful capability of forecasting the time series problem.
{
Γn = tna , tnd , Enreq
}
(2) The long short-term memory (LSTM) network is a variant of the RNN,
which can learn the long-term dependencies, and alleviate the problems
EVs staying in the charging station for a certain period of time can be caused by gradient disappearance and gradient explosion [33]. There­
regarded as a charging Task of the charging station. Define the Task set: fore, it can be well adapted to the forecast of traffic flows on special
VEC={Γ 1,…, Γ n,…, Γ N}, where N denotes the number of users in one dates, such as weekends and holidays. Na t, Nd t, Ns t, and ψ t show the
task, n is the index of the user. strong time-series correlate, therefore, based on past data, LSTM can be
The TOU electricity price at different time slots is defined as a vector: adopted to predict future EV charging demand information. Fig. 3 shows
e={e1,…,et,…,eT}, where T denotes the total time slot number, t is the the architecture of the LSTM unit. The compact forms of the equations
index of time slot. Therefore, the electricity purchase cost for the for the forward pass of an LSTM unit are [33,34]:
charging operator is:
( ) ft = σ (Wf ⋅[ht− Δt , Xt ] + bf ) (7)
G Pn,t = et Pn,t (3)
it = σ (Wi ⋅[ht− Δt , Xt ] + bi ) (8)
where Pn,t denotes the charged power of EV n at time slot t.
Ct = ft ∗ Ct− Δt + it ∗ tanh(WC ⋅[ht− Δt , Xt ] + bc ) (9)
2.2. Battery degradation characteristic
ot = σ (Wo ⋅[ht− Δt , Xt ] + bo ) (10)
From the perspective of users, EV users expect to complete the
ht = ot ∗ tanh (Ct ) (11)
charging requirements within the dwell time while increasing the bat­
tery life. Thus, the battery degradation model is established. The In the above equations, the forget gate’s activation vector is ft, the
LiFePO4 lithium-ion battery, which has more thermal and chemical
stability, has been widely used in a variety of EVs [30]. A degradation
cost model for LiFePO4 battery cells is developed in [31], which can be
expressed as:
( )
Φ Pn,t = at (Pn,t /Capn )2 + bt Pn,t /Capn + ct (4)

where parameters at, bt and ct are related to the battery

characteristics:

at = 106 tCcell β4 /(DVnom )

bt = 103 tCcell /(β2 − β6 Vnom ) (5)
2 3
ct = tDCcell Vnom /(β1 + β3 Vnom + β5 Vvom + β7 Vvom )

where Ccell is the price ($/Wh) of battery cell capacity; Vnom denotes Fig. 3. LSTM unit.

5
J. Liu et al. International Journal of Electrical Power and Energy Systems 141 (2022) 108218

input gate’s activation vector is it and the output gate’s activation vector (continued )
is ot; Ct represents the cell state vector; σ is the sigmoid function; W and b 22: The overall charging task is: VECτ = VECre es
τ ∪ VECτ where the users in task
are weight matrices and bias vector parameters which need to be learned are:Γn,τ ∈ VECτ
during training. The symbol * is the Hadamard multiplication. Xt and ht 23: Output: VECτ

denote the input and output vectors of the LSTM unit. In the operation
process of the proposed DICS algorithm, new data will be recorded to Let Na,re t, Nd,re t, and Ns,re t represent the number of real arrived,
update the LSTM module at each executing time slot. This can further departed, and stayed EVs at future time slot (t ∈ [τ + Δt, Tτ]); ψ re t
improve the accuracy of the forecast, and make the forecasting more denotes the total power of the real EVs at future time slot (t ∈ [τ + Δt,
adaptable to the rapid changes in the market share of EVs. Tτ]). Since the subsequent traffic is uncertain, the number of real arrived
Overall charging demand: Define τ as the current time slot, infor­ EV Na,re t at future time slot (t ∈ [τ + Δt,Tτ]) is 0, and the other infor­
mation about EVs arrived earlier than current time slot (ta n ≤ τ) is mation Nd,re t, Ns,re t, and ψ re t are determined by the current real
known, and some of their departure time later than current time slot (td charging task VECre τ. Therefore, at each time slot, the subsequent EVs
n > τ). In order to better schedule the charging scheme within the EVs’ information need to predict, and the established information set at time
dwell time, the information of EVs arrived after the current time slot (ta slot τ is called Estimated task VECes τ. Na,es t, Nd,es t, Ns,es t, and ψ es t
n > τ) needs to be predicted. For the EVs that stay in the station with represent the number of estimated arrived EVs, number of estimated
charging requirements at the current moment τ, their charging demand departed EVs, number of estimated stayed EVs, and total power of the
information is known. Thus, we define this part of EVs as Real EVs. estimated EVs at future time slot (t ∈ [τ + Δt,Tτ]). Therefore, by sub­
Define the set of real EVs tasks at the current time slot τ is: tracting the real EVs data from the predicted data, the estimated EVs,
{ ⃒ } can be obtained. Fig. 4 illustrates the relationship between the real and
re ⃒ a,re
VECre re d,re
τ = Γ1,τ , …, ΓNτre ,τ ⃒∀tn,τ ⩽τ⩽tn,τ (12) estimated EVs. Algorithm 1 shows the establishment of the overall
charging demand task VECτ at time slot τ, which is composed by VECre τ
where Nre τ denotes the number of EVs in VECre τ. Real EVs tasks at
and VECes τ. Nτ = N re τ + N es τ denotes the number of EVs in VECτ. The
the current time slot are composed of uncompleted charging tasks at last
arrival time of estimated EVs in VECes τ will not be earlier than τ + Δt, as
time slot VECre,un τ-Δt and EVs arriving at the station at current time
described ∀t a,es n,τ ≥ τ + Δt.
slot VECre,a τ, which is VECre τ = VECre,un τ-Δt ∪ VECre,a τ.
Since we need to formulate a charging scheduling strategy for the Fig. 5 shows the relationship between the established real and esti­
EVs currently staying in the station, the optimized time period is from mated charging tasks at different time slots. As the new time slot τ ar­
the current time slot τ to the time when the last EV. Define the interval to rives, the information of the EV entering the charging station at the
be predicted as [τ, Tτ], where Tτ indicates the latest departure time of the current time slot can be obtained, which is VECre τ. By adopting the
EVs in VECre τ, computed by: proposed Data-Driven EV Charging Demand Forecasting algorithm, we
{ ( )⃒ } can obtain the subsequent prediction results at the current moment and
Tτ = max tnd ⃒n ∈ Nτre (13) establish the Estimated task VECes τ. At the same time, the LSTM module
Let the previous information Na t, Nd t, Ns t and ψ t (∀t ≤ τ) be the will be upgraded as new data VECre τ becomes available, thus improving
input of the LSTM layer, the output of the LSTM layer is: the accuracy of subsequent predictions. Then the corresponding
{ ⃒ } charging scheduling optimization calculation is performed based on the
ht = Nta,pred , Ntd,pred , Nts,pred , ψ pred
t
⃒∀t ∈ [τ, Tτ ] (14) unfinished charging task at the last time slot VECre τ-Δt, new arrival task
VECre τ and predicted estimated task VECes τ. Then, at the next time slot
where Na,pred t, Nd,pred t, and Ns,pred t represent the predicted
τ + Δt, new EVs with charging task VECre τ + Δt will also enter the
number of arrived, departed and stayed EVs at future time slot (t ∈ [τ +
station. By repeating the above steps, the estimated task at time slot τ +
Δt,Tτ],τ + Δt means the next time slot after τ), respectively; ψ pred t is the
Δt can be predicated, and the optimal scheduling strategy for this time
predicted total energy demand at future time slot (t ∈ [τ + Δt,Tτ]).
slot can be obtained.
Algorithm 1. (Overall charging demand task establishment)
4. Optimization model and system scheduling
1: Input: VECre τ, VEC re τ− 1, Na,re t, Nd,re t, Ns,re t, ψ re t, τ, Tτ, ht
2: Initialization: n1=1, n2=1
The main target of the proposed charging scheduling scheme is to
3: for each t∈{τ+1,τ+2,…, Tτ}
reduce charging costs and increase users’ satisfaction. By introducing
4: while n1 ⩽Na,pred
t do
both real and estimated tasks into optimization, the charger operating
5: The arrival time of n1-th EV is set as tna,es =t
1
and charging power scheduling at the current time slot is more efficient
6: n1++
7: end while
8: Number of estimated EVs leaving at t: Nd,es
t = Nd,pred
t − Nd,re
t
9: if Nd,es
t > n1 − Nd,es
t

10: Nd,es
t = n1 − Nd,es
t

11: end if
12: while n2 ⩽Nd,es
t do
13: The arrival time of n1-th EV is set as tnd,es
2
=t
14: n2++
15: end while
( )/
16: Average power of estimated EVs at t: Pav,es
t = ψ pred
t − ψ re
t Ns,es
t

17: end for

18: for all the estimated EVs: n∈{1, 2,…,n1}
∑t=tnd,es
19: Power required by each estimated EV is Ereq,es
n = t=tna,es
ΔtPav,es
t

20: end for

21: Create the estimated EV users and the estimated charging task at current time slot
{ } { }
a,es d,es ini,es req,es req,es
τ :Γes
n,τ = tn,τ , tn,τ , Sn,τ , Sn,τ , En,τ and VECes es es
τ = Γ1,τ , …, ΓNes
τ ,τ

(continued on next column)

Fig. 4. Relationship between the real and estimated EVs.

6
J. Liu et al. International Journal of Electrical Power and Energy Systems 141 (2022) 108218

Fig. 5. Time-varying data prediction and charging task establishment process.

and economical. ( )2
tnd
∑
( )
G Pn,t = ω Enreq − ΔtPn,t (18)
4.1. Design of the optimization model
t=tna

Considering the charging network service capability, the charging where ω is the penalty factor which is larger enough. Therefore, by
status of each EV should be denoted. Therefore, the binary decision introducing this large enough penalty function, the charging station can
variable ζn,tthat indicates the connecting between chargers and EVs is be prioritized to complete the charging demand as the goal, and at the
defined as: same time, let the inability to fully complete the charging task be the
worst outcome. The larger difference between the obtained power and
ζn,t = {0, 1} , ∀n ∈ N, ∀t ∈ [τ, Tτ ] (15)
the required power, the greater the penalty function. Thus, the mathe­
Specifically, when ζn,t = 1, the EV n is connected to a charger and can matical program of the proposed model is:
be charged at the time slot t; otherwise when ζn,t = 0, the EV n is not [ ]
( ) ∑Nτ ∑
Tτ
( ) ( )
connected to any charger at time t, so it cannot get energy supply. Ω Pn,t , ζn,t = min F Pn,t , ζn,t + ζn,t G Pn,t (19)
By considering the charging cost, battery degradation cost, the
Pn,t ,ζm,t
n=1 t=τ

comprehensive charging cost of the n-th EV at time t can be established subject to:
as: (4), (16), (18).
( ) ( ( ))
F Pn,t , ζn,t = ζn,t et Pn,t + Φ Pn,t (16) ∑Nτ
ζ ⩽M , ∀t ∈ [τ, Tτ ] (20a)
n=1 n,t
In order to ensure user satisfaction, EVs need to be fully charged
[ ]
before departure. The following constraint is: ∈ tna , tnd , ∀n ∈ N
ζn,t = 0 , if ∀t ∕ (20b)
∑tnd
ΔtPn,t = Enreq , ∀n (17) Pn,t ⩽Pmax , ∀n ∈ N, ∀t ∈ [τ, Tτ ] (20c)
t=tna

However, due to the traffic flow changes in real-time, it is not Pn,t = 0 , if ζn,t = 0, ∀n ∈ N, ∀t ∈ [τ, Tτ ] (20d)
possible to guarantee that all EVs can be fully charged with limited
chargers. From the user’s perspective, EVs are expected to complete the ∑tnd
ΔtPn,t ⩽Enreq , ∀n ∈ N (20e)
charging requirement as much as possible before departure. Thus, the t=tna

penalty function method is introduced to evaluate user satisfaction. The

In the established model (19)–(20), the object is to minimize the
penalty function of n-th EV is:
comprehensive charging cost including the cost of TOU electricity price,

7
J. Liu et al. International Journal of Electrical Power and Energy Systems 141 (2022) 108218

the cost of battery degradation, and the penalty part for not meeting the
2at Δt ( ) ( )
user’s charging demand. The decision variable is set as the charging λ1 = + 2ωΔt tnd − tna + 2ωΔt tnd − tna − 1
Capn
status ζn,t and the charging power Pn,t. Constraint (20a) expresses the (25)
maximum M chargers allowed to work at the same time, M represents λ2 = … = λt =
2at Δt ( )
+ 2Δtω tnd − tna − 2Δtω
the number of chargers in the charging station. Constraint (20b) means Capn
EVs cannot be charged when they were not at the charging station. The It can be found that ω is the penalty factor which is positive and
power provided by each charger at a time slot has an upper limit, which much larger enough, the departure time of an EV is always larger than its
is represented by constraint (20c). Constraint (20d) indicates that when arrival time, at and Capn are also positive value. Thus all the eigenvalues
the charging state ζn,t is 0, the EV charging power Pn,t is also 0. Since we of the Hessian matrix are equal or larger than zero. Thus, for all vector x
have introduced the penalty function (18) to indicate the negative in Rn, we have: H positive semi-definite ⇔ xTHx≥0. Constraints (20c)-
impact of incomplete EV charging, the constrain (17) is converted into (20e) are also linear constraints. Thus, the objective function is convex
constraint (20e). Constraint (20e) means the obtained energy of one EV when the binary variable ζn,t is fixed.
cannot exceed its required power.
Algorithm 2. (Heuristic algorithm)
4.2. Solving technique
1: Set: Generation number g, Population size μ, Crossover probability λ, Mutation
probability δ.
Since the established model is based on data prediction, there is no 2: Input: VECτ
exact optimal solution. And the optimization model is mixed-integer 3: Initialization: Randomly create μ different individuals ζ μ under constraints
nonlinear programming (MINLP) that includes binary variable ζn,t and (20a)-(20b), where ζ μ¼{ζ1,τ,…, ζn,t}.
natural number variable Pn,t. Since we cannot, in general, construct 4: while termination criteria, g>gmax, not fulfilled, do
5: Fitness value: with the fixed ζ μ, use a convex optimization method to find the
causal optimal scheduling policies [10,35], the corresponding subopti­ solution of lower model (objective (19) constraints (20c)-(20e)) for μ different
mal heuristic solving algorithm is developed in this paper. According to individuals, record the solved decision variable Pμ={P1,τ,…,Pn,t}. Ωμ(Pμ,ζ μ) is the
convex optimization theory we have: fitness value.
6: Best individual record: record the individual with the smallest Ωμ. if Ωμ<Ωbest
Proposition 1. when the binary variable ζn,t is fixed, the optimization 7: Ωbest =Ωμ, Pbest =Pμ, ζ best=ζ μ
model with only natural number variables is convex. 8: end if
9: Select: select individuals to produce offspring while individuals with smaller Ω
Proof 1. Let ζn,t be the fixed constants, compute the partial derivative of the values are easier to be selected.
function (19) with respect to variable Pn,t: 10: Offspring produce: the selected adjacent individuals have λ rate to exchange
chromosomes, and the exchange position is random. Produce the offspring
∂Ω 2at bt individual ζ μ.
= ζn,t et + ΔtPn,t + 11: Mutation: the offspring individual ζ μ has δ rate to mutate.
∂Pn,t Capn Capn
12: end while
[ ] [ ]
∑tnd ∑tnd (21) 13: Output: Ωτ=Ωbest, Pτ=Pbest, ζ τ=ζ best
− 2ω Enreq − t=ta
ΔtPn,t − ⋯ − 2ω Enreq − t=t a
ΔtPn,t
n n

⏟̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅⏞⏞̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅⏟ Therefore, by dividing the optimization model into two layers, the
number depends on: tnd − tna
search range of the heuristic algorithm is reduced significantly, the
Continue to find the second partial derivative with respect to vari­ computational efficiency is improved and more feasible solutions are
able Pn,t, we have: obtained. The upper model is for solving binary variables ζn,t, where the
objective function is (19) and constraints are (20a)-(20b). The lower
∂Ω2 2at Δt model is set ζn,t as a fixed value and takes an objective function as (19)
= + 2ωΔt(tnd − tna ) (22) and constraints as (20c)-(20e) to solve charging power Pn,t. Algorithm 2
∂P2n,t Capn
shows the solving method of the charging scheduling model based on a
Then find the partial derivative based on (21) with respect to vari­ genetic algorithm [28].
able Pn,t’(t’∕
=t), we have:

∂Ω2 4.3. Operation of the charging scheduling system

(23)
′
= 2Δtω , t ∕=t
∂Pn,t ∂Pn,t′
Because the real and estimated EV tasks, VECre τ and VECes τ,
Therefore, the Hessian Matrix of the objective function is obtained: change in time series, the overall charging task VECτ and corresponding
⎡ ⎤ results from Algorithm 2 should be updated at a different time τ. In the
∂Ω2 ∂Ω2
⎢ ∂P ⋯ optimization model (19)-(20), the t ∈ [τ,Tτ], but the results Pn,t and ζn,t,
2
∂Pn,1 ∂Pn,t ⎥
⎢
⎢
n,1 ⎥
⎥ when t ≥ τ + Δt, will not be adopted as the scheduling scheme. Because
⎢ ⎥
H=⎢ ⋮ ⋱ ⋮ ⎥ the estimated EVs are included at t ∈ [τ + Δt,Tτ], which does not actually
⎢ ⎥
⎣ ∂Ω 2
∂Ω 2 ⎦ exist in the reality. Therefore, we only keep Pn,τ and ζn,τ (n ∈ Nτ) as the
⋯
∂Pn,1 ∂Pn,t 2
∂Pn,t charging scheme at the current time slot τ.
⎡
2at Δt ( d ) ⎤ The complete charging scheduling system operation process is shown
in Fig. 6. Based on the traffic data, the LSTM algorithm is applied to
a
⎢ Capn + 2ωΔt tn − tn ⋯ 2Δtω ⎥
⎢
⎢
⎥
⎥ predict the subsequent traffic data. Combining with the current EV data
=⎢ ⋮ ⋱ ⋮ ⎥ (24)
⎢ ⎥ VECre τ, the forecasted EV flow data ht, and the unfinished task VECre
⎣ 2at Δt ( ) ⎦
2Δtω ⋯ d a
+ 2ωΔt tn − tn τ-Δt, the overall charging task at current time slot VECτ is established by
Capn adopting Algorithm 1. Then, with the overall charging task VECτ, the
Solving the eigenvalues of the Hessian matrix: optimization model (19)-(20) is formed and Algorithm 2 is introduced as
the solving technique. By considering TOU price, battery degradation,
and user’s satisfaction, with the proposed Algorithm 2, the model which
takes Pn,t, and ζn,t as the decision variables to minimize the overall
charging cost is computed. The solved results are matrixes Pτ[Nτ×(Tτ −
τ + 1)] and ζ τ[Nτ×(Tτ − τ + 1)] including the charging scheme of Nτ EVs

8
J. Liu et al. International Journal of Electrical Power and Energy Systems 141 (2022) 108218

Fig. 6. Charging scheduling system operation process.

from time τ to Tτ. All the EVs that receive energy replenishment at the (continued )
current time slot τ are Real EVs. Thereby, the first column of charging 3: Preq
n = Preq
n − Pn,t
status ζ in τ and charging power P in τ are taken as the input signal, 4: If Preq
n ⩽0
where ζ in τ={ζ1,τ,…,ζn,τ} and P in τ={P1,τ,…,Pn,τ}. ζ in τ and P in τ are
5: Remove Γ n,τ from VECτ.
regarded as the control signal at time slot τ to guide the charging station 6: end if
operation. Then, the current charging task VECτ is updated, that is, the 7: end for
user’s charging demand is updated, and users who have completed the 8: Output: VECτ
charging demand will be removed from the charging task set (Algorithm
3). After that, the updated charging task at τ is retained until the next
time slot τ + Δt for establishing the new charging task VECτ+Δt. Repeat
the above steps continuously to realize the sequential operation of the 5. Case study
charging station.
In the case study, the number of chargers in the charging station is set
Algorithm 3. (Charging task update)
to M = 20, all EVs and chargers are the same types and the maximum
1: Input: VECτ charging power per hour provided by each charger is Pmax = 45 kW. The
2: For each n∈{1, 2,…,Nτ} sequential EV charging requirements follow theprobability density
(continued on next column) function mentioned above, which reflect the charging requirements in
the real world. The penalty parameter ω is set to a large positive value

9
J. Liu et al. International Journal of Electrical Power and Energy Systems 141 (2022) 108218

109. The time slot duration Δt is set as one hour. The TOU price is
following the German electricity spot prices [36]. The traffic flow is the
topical scenario of workplace that higher on workdays and lower on
weekends. The initial SOC and required energy value are generated by
following the probability density function. In the case study, the average
number of EVs charging in a workday and a rest day are set as 100 and
20 respectively.
Fig. 7 illustrates the comparison of forecasted data under LSTM and
real data, including the number of dwell EVs, number of arriving EVs,
and expected power at each time slot. According to the introduced Al­
gorithm 1, the set of estimated charging tasks can be obtained. The
estimated charging task VECes τ will be established at every operating
time slot based on historical data and current station information. So the
estimated charging task established from the perspective of the different
time slots is different. Thereby, an example that the established subse­
quent estimated charging task from the perspective of 7:00 (VECes 7) is
presented in Fig. 8. Because the latest real EV plans to leave at 19:00
(which means in VECre 7, max(td n) = 19:00). So the estimated EVs task
VECes 7 we need to predict is from 8:00 to 19:00. Each square represents
the state of an EV at one time slot, where the row represents the index of
EV and the column denotes the time slot. The blue color represents the Fig. 8. Forecast the charging demand of the subsequent EV from the perspec­
EV staying in the charging station, and the deeper the blue the greater tive of 7:00.
the charging power required. It can be seen that the number of estimated
EVs is close to the real EVs, and overall, the trend of demand energy and EV. The Convectional forecasting optimization (CFO) scheduling
dwell time is also similar. The establishment of the future estimated task method is taken as the comparison. The CFO method only considers the
is only to guide the scheduling at the current time and leave a certain scheduling strategy in a centralized manner. The specific energy de­
margin for the subsequent vehicles. Therefore, it is not essential to know mands and battery health of each EV, and the specific operating scheme
the accurate charging demand information of each specific subsequent of each charger are ignored. Therefore, CFO has fewer constraints and
EV, and certain differences are allowed. can provide better scheduling results from the perspective of the
Compared with the existing scheduling methods that consider charging station. Fig. 9 shows the overall charging power and the
charging demand forecasting, the proposed data-based intelligent accumulated charging cost of one charging station under DICS and CFO.
charging scheduling (DICS) method can provide a detailed charger It can be found in Fig. 9 (a), even though more individual factors are
operating scheme and track the charging demands of each individual taken into consideration, the proposed DICS method is the same as the

Fig. 7. Traffic flow forecast.

10
J. Liu et al.
Fig. 9. Comparisons between the proposed DICS and CFO (a) charging power (b) accumulated charging cost.
CFO method, which avoids high-power charging when the TOU price is
high and charging as much as possible when the TOU price is low. Ac­
cording to Fig. 9 (b), the accumulated charging cost under the proposed
DICS is also very close to the ideal CFO situation. Meanwhile, the DICS
method can also provide a detailed operating scheme with limited
charging facilities, which cannot be provided by the CFO.
At every time slot, the proposed DICS approach will perform the
optimization process. The results of the optimization process at each
time slot can not only provide the charging scheme for the EVs staying in
the station at the current time slot but also predict the scheduling
scheme of subsequent EVs from the perspective of the current. In order
to better show the optimal results of the proposed DICS algorithm at
each time slot, the detailed power scheduling and charger operating
heatmap of each individual EV at 7:00 and 8:00 are shown in Fig. 10.
The gray box indicates that the EV is not in the station at this time and
cannot be charged. The other boxes indicate that the EV right now stays
in the charging station. The white to red gradient color represents the
amount of provided charging power. Meanwhile, the box with the green
dot denotes this EV is connecting to a charger. The limitation of chargers
is considered, the maximum number of chargers that can be activated at
each time slot is M, which is set to 20 in the simulation, so each column
has a maximum of 20 green dots. In Fig. 10 (a), the EVs that stay in the
charging station at 7:00 are all real EVs, and the established estimated
EVs will “arrive” at the charging station after 7:00. So the leftmost
column represents the real charged energy at 7:00. The second to last
columns represent the predicted charging energy of both real EVs and
estimated EVs. The last column represents the time when the last real EV
left. The marked Leave represents EVs that completed charging or de­
parture at this time slot and the marked Arrive represents EVs arriving at
this time slot. Similarly, Fig. 10 (b) shows the results from the
perspective of 8:00.
Comparing Fig. 10 (a) and Fig. 10 (b), EVs that completed their
charging task at 7:00 will not continue to occupy the charging resources
at 8:00, while EVs that have not completed charging will perform
11
International Journal of Electrical Power and Energy Systems 141 (2022) 108218
charging scheduling optimization together with newly arrived EVs at
8:00. If the information of subsequent EVs is ignored (scheduling only
considers the real EV part), from the perspective of the current time slot,
the number of EVs in the station will become less and less in the future,
resulting in the wrong judgment that there will be enough chargers in
the future, and then formulating unreasonable charging strategies. As
new EVs continue to enter the station at subsequent moments, the
charging incomplete rate will increase. With the introduction of predi­
cated estimated EVs task, the charging scheduling is more reasonable. In
addition, the “arrival” time of the estimated task also ensures that there
will be no invalid result output at the current time (send the operating
scheme to an estimated EV). Notice that the prediction for the estimated
EVs only plays a role in assisting the charging optimization for the real
EVs at the current time slot. The current estimated EVs will be cleared
and the new estimated task will be created at the next moment. At each
time slot, we can get the detailed scheduling heatmap results as in
Fig. 10 (a) or Fig. 10 (b). The leftmost column of the obtained heatmap is
used as the current charging station scheduling plan, and the remaining
columns are used as the prediction.
At every time slot when an EV dwells in the charging station, the
DICS will formulate the specific scheduling scheme and predict the
subsequent scheduling scheme for the EV, by considering factors such as
TOU price, battery degradation, and charger assignment. Fig. 11 shows
in detail the process of obtaining charging energy for a specific EV. This
EV arrived at 7 and left at 16, with 47 kWh energy requirement. Each
sub-picture in Fig. 11 represents a different time slot, and only the time
slot when the EV is connected to a charger is given here. By the DICS
approach, this EV was connected to a charger at 7:00, 13:00, 14:00, and
15:00 with charged energy 14.9 kWh, 8kWh, 9.9kWh, and 14.2kWh
respectively. When at a certain time slot, the DICS will charge specific
energy to this EV and then provide the predicted charging scheme for
subsequent moments. However, as the real arrival EV data differs
slightly from the predicted EV data, the real charging power at this time
slot will be readjusted. For example, shown in Fig. 11 (a), from the
J. Liu et al. International Journal of Electrical Power and Energy Systems 141 (2022) 108218

Fig. 10. Detailed power scheduling and charger operating scheme heatmap of each individual EV at (a) 7:00 (b) 8:00.

perspective of 7:00, the charging station provided 14.7kWh energy to (Ereq

∑tdn req
n − tan Pn,t )/En . To illustrate the performance of the proposed al­
this EV and predicted the future charging power of this EV. Fig. 11 (b)
gorithm, the key performance indexes, including provided total energy,
shows the charging power and predicted power at time slot 13:00. It can
total charging cost, number of dissatisfied users, and average dissatis­
be seen that the charged power at this time is different from the pre­
faction degree, under different scheduling methods are demonstrated in
dicted from the perspective of 7:00. This is because the algorithm adjusts
Table 1. Obviously, in the designed case, the FCFS guarantees all EVs
the charging power according to the actual arrival of the vehicle.
complete their charging requirements. However, because the FCFS does
Similarly, Fig. 11 (c) shows the power charged at 14:00 and the subse­
not perform charging power scheduling based on TOU price, the econ­
quent charging prediction. Fig. 11 (d) shows the power charged at
omy is the lowest. By adopting SWP, the total charging cost is reduced,
15:00. Since the charging of the EV is just completed at this time, there is
but due to the lack of subsequent EV information, the number of not
no need to charge at the next moment. Chargers responsible for this EV
fully charged EVs is the most, and the charge rate of these EVs is also
can be deployed to serve other EVs when they are not connected to this
low. SWR is the ideal situation, thereby, it can ensure that most EVs can
EV, thereby, the charging efficiency is improved along with the charger
complete the charging requirements while improving the charging
utilization rate.
economy. Considering the realistic scheduling, the results of perfor­
Because the CFO method is a centralized scheduling method and
mance indexes under DICS are as good as the ideal situation. Therefore,
ignores the operating of specific chargers and EVs, it does not have the
by applying the proposed DICS, with practical feasibility, the charging
conditions for direct comparison with DICS considering specific limited
economy is improved and user satisfaction is also guaranteed.
charging facilities. In order to demonstrate the performance of our
The service capacity is restricted by the number of chargers, thus the
proposed DICS, first come first serve scheduling (FCFS), scheduling
Monte Carlo simulation is taken for further evaluation. Set the daily
without prediction (SWP) and scheduling with real data (SWR) are taken
traffic volume to 100 and change the number of chargers. After 1000
for comparisons. Among them, SWP only considers the current EV data
times Monte Carlo simulations, the key performance indexes are shown
for optimization and does not establish the estimated EV task. SWR is the
in Fig. 12. From Fig. 12 (a), the smaller number of chargers provides
ideal condition that real EV data is used to replace the predicted esti­
lower total charging power. The total energy that the charging station
mated EV task. In the FCFS, EVs tend to finish charging as fast as
can provide under SWP is the least. The DICS and SWR can provide as
possible, directly charging with the maximum power and disconnecting
much power as FCFS. From the perspective of charging costs, shown in
the charger immediately after the charging is completed.
Fig. 12 (b), FCFS is the uneconomic way to operate charging stations.
To evaluate the satisfaction of individual EV users, the EV users who
The proposed DICS method greatly reduces the charging cost, which is as
have not completed their charging requirements are dissatisfied users,
∑td good as the ideal state SWR. According to Fig. 12 (c) and (d), SWP is
which is tnna Pn,t ⩽Ereq
n . For the dissatisfied users, the rate of incomplete more economical when the charger number is small, because it com­
charging indicates the dissatisfaction degree, which is pletes less charging demand and leads to a higher dissatisfaction degree.

12
J. Liu et al. International Journal of Electrical Power and Energy Systems 141 (2022) 108218

Fig. 11. The actual charging power and the predicted charging power at different times of one EV (a) 7:00 (b) 13:00 (c) 14:00 (d) 15:00.

However, when the charger number increases, SWP costs more than the
Table 1
proposed DICS. FCFS is the best scheduling method to ensure user
Key performance indexes under different scheduling methods.
satisfaction, but it is also the most expensive to charge. In comparison,
DICS FCFS SWP SWR the proposed DICS, which is remarkably close to the ideal state SWR, not
Provided total energy (kWh) 24,707 24,741 24,244 24,642 only reduces charging costs but also keeps users’ dissatisfaction within
Total charging cost (€) 7638 8847 7951 7615 an acceptable range.
Number of dissatisfied users 3 0 28 4
Fig. 13 shows the boxplot of different indexes under different
Average dissatisfaction degree (%) 19.4 0 34.3 21.1
numbers of chargers according to Monte Carlo simulation when DICS is
adopted. It can be seen that as the number of chargers increases, the

13
J. Liu et al.
Fig. 12. Mean value of the key performance indexes under different number of chargers (a) Total provided charging power (b) Total charging cots (c) Number of
dissatisfied users (d) Dissatisfaction degree.
Fig. 13. Monte Carlo simulation results of different indexes under different numbers of chargers when adopted DICS (a) Total provided charging power (b) Total
charging cots (c) Number of dissatisfied users (d) Dissatisfaction degree.
results of charging power, charging cost, and dissatisfied users become
more concentrated. Because the greater number of chargers brings
greater scheduling flexibility, which improves the economy while
making the results more concentrated. By contrast, also the average
dissatisfaction degree reduced as the chargers increased, the Monte
Carlo simulation results were still scattered. Because when the number
of chargers is large, the number of dissatisfied users is small and arrived
EVs are random, so the Monte Carlo results are often determined by a
few special EVs (such as only one EV in the entire system that does not
complete 80% of its charging requirement, so the dissatisfaction degree
in this time is 80%). Therefore, we believe these results are acceptable.
Since the battery degradation factor is taken into account in the
proposed algorithm, the battery life performance improvement under
Monte Carlo simulation is discussed. The scheduling without battery
degradation (SWBD) method has been discussed in some previous works
Fig. 14. Battery and charging cost-saving rate of the proposed DICS and the
SWB, N = 100.
14
International Journal of Electrical Power and Energy Systems 141 (2022) 108218
[1,37]. Thus, the battery cost-saving rate of SWBD and DICS compared
to FCFS is shown in Fig. 14, together with the charging cost-saving rate.
Obviously, with the number of chargers increasing, both two methods
are more economical than FCFS. The charging cost under the SWBD is
slightly less than that under the DICS, but there is nearly no positive
impact on the EV’s battery life. Because SWBD charges as much energy
as possible when the TOU price is the lowest, which is not beneficial to
the battery. In contrast, the DICS method can balance battery degrada­
tion and charging cost, and make a positive contribution to both battery
performance and charging economy.
6. Conclusion
To improve the scheduling operation economy of charging stations
with limited charging facilities under the uncertainty of charging de­
mand, a data-based intelligent charging scheduling method is proposed
in this paper. Based on historical traffic data, a neural network-based
algorithm is adopted to predict the follow-up traffic flow and establish
the estimated charging demand task. Through comprehensive consid­
eration of estimated and real tasks, the optimization model considering
charging cost, battery degradation, users’ satisfaction, and limitation of
charging facilities is established. With the introduced solving algorithm,
the charging scheme is optimized for guiding chargers’ operating. Real-
time scheduling of charging stations is achieved by updating the neural
network and charging task status. The simulation results verify the
effectiveness of the proposed scheduling method. Furthermore, four key
performance indexes are introduced and compared among the proposed
DICS method, FCFS, SWP, and SWR methods. In addition, the battery
performance improvement compared with the SWBD is also illustrated.
According to Monte Carlo simulation results, with the proposed DICS
method the charging operators can not only achieve better economic
performance as much as possible but also improve service completion
J. Liu et al.
rate through predictive algorithms and user satisfaction, which is better
than other existing algorithms. The proposed DICS method is the
scheduling approach for a single charging station. Therefore, in the
future, the study on coordinated charging scheduling of multiple
charging stations in the transportation and distribution network inter­
action system will be carried on. EV charging path planning based on
traffic flow data will be considered, and the scheduling method that can
reduce the impact of EV charging on the power grid will be formulated.
CRediT authorship contribution statement
Jiayan Liu: Conceptualization, Methodology, Software, Visualiza­
tion, Writing – original draft. Gang Lin: Validation, Resources, Writing –
review & editing. Christian Rehtanz: Supervision, Writing – review &
editing. Sunhua Huang: Investigation. Yang Zhou: Data curation.
Yong Li: Writing – review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Acknowledgment
This work is based upon work in the project KonVeEn (Grant No.
03EK3057A, supported by the German Federal Ministry of Education
and Research), in part by the “111” Project of China under Grant
B17016, and supported in part by China Scholarship Council (No.
201806130201).
References
[1] Zhang Y, You P, Cai L. Optimal Charging Scheduling by Pricing for EV Charging
Station With Dual Charging Modes. IEEE Trans Intell Transp Syst Sept. 2019;20(9):
3386–96.
[2] Liu J, Peper J, Lin G, Zhou Y, Awasthi S, Li Y, et al. A planning strategy considering
multiple factors for electric vehicle charging stations along German motorways. Int
J Electr Power Energy Syst Jan. 2021;124:106379.
[3] Banol Arias N, Tabares A, Franco JF, Lavorato M, Romero R. Robust Joint
Expansion Planning of Electrical Distribution Systems and EV Charging Stations.
IEEE Trans Sustainable Energy 2018;9(2):884–94.
[4] Xie R, Wei W, Khodayar ME, Wang J, Mei S. Planning Fully Renewable Powered
Charging Stations on Highways: A Data-Driven Robust Optimization Approach.
IEEE Trans Transp Electrif 2018;4(3):817–30.
[5] “Sinocharged: The bright future of China’s electric vehicle market,” [Online].
Available: https://assets.kpmg/content/dam/kpmg/cn/pdf/en/2021/01/2020-ch
ina-leading-autotech-50.pdf. [Accessed 5 7 2021].
[6] “China Charging Infrastructure Trend Report (2020 to 2025),” [Online]. Available:
https://pdf.dfcfw.com/pdf/H3_AP202008311404859896_1.pdf?1598997513000.
pdf. [Accessed 5 7 2021].
[7] Zhang H, Hu Z, Xu Z, Song Y. Optimal Planning of PEV Charging Station With
Single Output Multiple Cables Charging Spots. IEEE Trans Smart Grid Sept. 2019;8
(5):2119–28.
[8] Biondi E, Boldrini C. and Bruno R. “The impact of regulated electric fleets on the
power grid: The car sharing case,” in IEEE 2nd International Forum on Research and
Technologies for Society and Industry Leveraging a better tomorrow (RTSI), Bologna,
Italy, 2016.
[9] Chen H, Hu Z, Luo H, Qin J, Rajagopal R, Zhang H. Design and Planning of a
Multiple-Charger Multiple-Port Charging System for PEV Charging Station. IEEE
Trans Smart Grid 2019;10(1):173–83.
[10] Wei Z, Li Y, Zhang Y, Cai L. Intelligent Parking Garage EV Charging Scheduling
Considering Battery Charging Characteristic. IEEE Trans Ind Electron Mar. 2018;65
(3):2806–16.
15
International Journal of Electrical Power and Energy Systems 141 (2022) 108218
[11] Song Y, Zheng Y, Hill DJ. Optimal Scheduling for EV Charging Stations in
Distribution Networks: A Convexified Model. IEEE Trans Power Syst 2017;32(2):
1574–5.
[12] Gupta V, Konda SR, Kumar R, Panigrahi BK. Multiaggregator Collaborative Electric
Vehicle Charge Scheduling Under Variable Energy Purchase and EV Cancelation
Events. IEEE Trans Ind Inf 2018;14(7):2894–902.
[13] Cai H, Chen Q, Guan Z, Huang J. Day-ahead optimal charging/discharging
scheduling for electric vehicles in microgrids. Protection and Control of Modern
Power Systems 2018;3:9.
[14] Farzin H, Fotuhi-Firuzabad M, Moeini-Aghtaie M. A Practical Scheme to Involve
Degradation Cost of Lithium-Ion Batteries in Vehicle-to-Grid Applications. IEEE
Trans Sustainable Energy Oct. 2016;7(4):1730–8.
[15] Wei Z, Li Y, Cai L. Electric Vehicle Charging Scheme for a Park-and-Charge System
Considering Battery Degradation Costs. IEEE Trans Intell Veh 2018;3(3):361–73.
[16] Hoke A, Brissette A, Smith K, Pratt A, Maksimovic D. Accounting for Lithium-Ion
Battery Degradation in Electric Vehicle Charging Optimization. IEEE Journal of
Emerging and Selected Topics in Power Electronics 2014;2(3):691–700.
[17] Alinia B, Hajiesmaili MH, Crespi N. Online EV Charging Scheduling With On-
Arrival Commitment. IEEE Trans Intell Transp Syst Dec. 2019;20(12):4524–37.
[18] Hou L, Wang C, Yan J. Bidding for Preferred Timing: An Auction Design for Electric
Vehicle Charging Station Scheduling. IEEE Trans Intell Transp Syst Aug. 2020;21
(8):3332–43.
[19] Ghosh A, Aggarwal V. Control of Charging of Electric Vehicles Through Menu-
Based Pricing. IEEE Trans Smart Grid Nov. 2018;9(6):5918–29.
[20] Zeng B, Dong H, Xu F, Zeng M. Bilevel Programming Approach for Optimal
Planning Design of EV Charging Station. IEEE Trans Ind Appl Feb. 2020;56(3):
2314–23.
[21] Pflaum P, Alamir M. and Lamoudi MY. “Probabilistic Energy Management Strategy
for EV Charging Stations Using Randomized Algorithms,” IEEE Transactions on
Control Systems Technology, vol. 26, no. 3, pp. 1099-1106, May. 2018.
[22] Şengör İ, Erdinç O, Yener B, Taşcıkaraoğlu A, Catalão JPS. Optimal Energy
Management of EV Parking Lots Under Peak Load Reduction Based DR Programs
Considering Uncertainty. IEEE Trans Sustainable Energy Jul. 2019;10(3):1034–43.
[23] Fallah-Mehrjardi O, Yaghmaee MH, Leon-Garcia A. Charge Scheduling of Electric
Vehicles in Smart Parking-Lot Under Future Demands Uncertainty. IEEE Trans
Smart Grid Nov. 2020;11(6):4949–59.
[24] Zhang T, Chen W, Han Z, Cao Z. Charging Scheduling of Electric Vehicles With
Local Renewable Energy Under Uncertain Electric Vehicle Arrival and Grid Power
Price. IEEE Trans Veh Technol Jul. 2014;63(6):2600–12.
[25] Jahangir H, Tayarani H, Ahmadian A, Golkar MA, Miret J, Tayarani M, et al.
Charging demand of Plug-in Electric Vehicles: Forecasting travel behavior based on
a novel Rough Artificial Neural Network approach. J Cleaner Prod Aug. 2019;229:
1029–44.
[26] Wan Z, Li H, He H, Prokhorov D. Model-Free Real-Time EV Charging Scheduling
Based on Deep Reinforcement Learning. IEEE Trans Smart Grid Sept. 2019;10(5):
5246–57.
[27] Li H, Wan Z. and He H. “Constrained EV Charging Scheduling Based on Safe Deep
Reinforcement Learning,” IEEE Transactions on Smart Grid, vol. 11, no. 3, pp. 2427-
2439, May. 2020.
[28] Liu J, Lin G, Huang S, Zhou Y, Li Y, Rehtanz C. Optimal EV Charging Scheduling by
Considering the Limited Number of Chargers. IEEE Trans Transp Electrif Sept.
2021;7(3):1112–22.
[29] Soares FJ, Almeida PMR, Lopes JAP. Quasi-real-time management of Electric
Vehicles charging. Electr Power Syst Res 2014;108:293–303.
[30] Tan X, Qu G, Sun B, Li N, Tsang DHK. Optimal Scheduling of Battery Charging
Station Serving Electric Vehicles Based on Battery Swapping. IEEE Trans Smart
Grid Mar. 2019;10(2):1372–84.
[31] Ma Z, Zou S, Ran L, Shi X, Hiskens IA. Efficient decentralized coordination of large-
scale plug-in electric vehicle charging. Automatica 2016;69:35–47.
[32] Forman J. Stein J. and Fathy H. “Optimization of dynamic battery paramter
characterization experiments via differential evolution,” in American Control
Conference, Washington, DC, USA, Jun. 2013.
[33] Gers F. Schmidhuber J. and Cummins F. “Learning to forget: continual prediction
with lstm,” in Ninth International Conference on Artificial Neural Networks ICANN
99., Edinburgh, UK, 1999.
[34] Mackenzie J, Roddick JF. and Zito R. “An Evaluation of HTM and LSTM for Short-
Term Arterial Traffic Flow Prediction,” IEEE Transactions on Intelligent
Transportation Systems, vol. 20, no. 5, pp. 1847-1857, May. 2019.
[35] Subramanian A, Garcia M, Domínguez-García A, Callaway D, Poolla K. and Varaiya
P. “Real-time scheduling of deferrable electric loads,” in American Control
Conference (ACC), Montreal, QC, Canada, 2012.
[36] “smard,” 12 2020. [Online]. Available: https://www.smard.de/home/strommarkt-
aktuell/alle-artikel.
[37] Cheng X, Hu X, Yang L, Husain I, Inoue K, Krein P, et al. Electrified Vehicles and
the Smart Grid: The ITS Perspective. IEEE Trans Intell Transp Syst Aug. 2014;15(4):
1388–404.