Available online at www.sciencedirect.

com

ScienceDirect
Energy Procedia 73 (2015) 173 – 181

9th International Renewable Energy Storage Conference, IRES 2015

Online optimal charging strategy for Electric Vehicles

Chenjie Maa*, Juha Rautiainena, Dirk Dahlhausb, Akhilesh Lakshmana,
J.-Christian Toebermanna and Martin Brauna,c
a
Fraunhofer Institute for Wind Energy and Energy System Technology (IWES), Kassel, Germany
b
Communications Laboratory, University of Kassel, Kassel, Germany
c
Energy Management and Power System Operation, University of Kassel, Kassel, Germany

Abstract

This work investigates scheduling strategies for charging electric vehicles (EVs) in distribution grids. Our proposed scheduling
strategy is realized as a moving window optimization scheme. By considering load and price forecasting and EVs’ power
demand, this strategy optimizes the charging costs of an EV fleet with consideration of grid constraints. Its benefit is first
demonstrated through simulation studies on workday and weekend scenarios. Furthermore, we discuss the impact of scheduling
windows on the optimization result. Last, the impact of forecasting errors of energy prices on the optimization performance is
quantitatively analyzed. The results show that the proposed strategy can optimally plan EV charging schedules with consideration
of online information and uncertainties.

© 2015 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
© 2015 The Authors. Published by Elsevier Ltd.
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of EUROSOLAR - The European Association for Renewable Energy.
Peer-review under responsibility of EUROSOLAR - The European Association for Renewable Energy
Keywords:electric vehicle, online optimization, charging, scheduling.

1. Introduction

Electric mobility has the potential to facilitate the future transportation demand with high energy efficiency and
low green-house gas emissions. Potential impacts on distribution systems by charging of EVs have been investigated
by recent publications. Depending on the applied charger technology, congestions may already exist in present

*Corresponding author. Tel.: +49 561 7294 106; fax: +49 561 7294 200.
E-mail address:chenjie.ma@iwes.fraunhofer.de

1876-6102 © 2015 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of EUROSOLAR - The European Association for Renewable Energy
doi:10.1016/j.egypro.2015.07.667
174 Chenjie Ma et al. / Energy Procedia 73 (2015) 173 – 181

distribution grids even at a low penetration rate of EVs [1]. Other problems such as the increase in voltage variations
and system losses are also pointed out in [1~3].
To mitigate these risks, some studies proposed intelligent scheduling concepts for charging of EVs. As multiple
stakeholders with different interests are involved, this scheduling problem expresses a highly dynamic, multi-
objective complexity. EV owners naturally want to charge their EVs as fast as possible, while commercial
aggregators or charging service providers have another clear objective to minimize the energy procurement cost and
accordingly maximize their profit. Providing ancillary services to the grid may be a viable option, only if a proper
incentive scheme or a market exists. From the perspective of a distribution system operator (DSO), the paramount
task is to operate the grids efficiently and within technical limits. In such a complex context, where the interests of
all stakeholders must be considered, a technical or economical optimum is difficult to achieve. In previous studies,
different aspects in representing these entities as direct objectives or as limitations are discussed.
Starting from the aspect of a system operator, strategies to maximize the delivered power to charge EV fleets are
proposed by [4] and [5]. These authors provide solutions with pure technical considerations, where power rate of
chargers, current and voltage limits of the grid and power delivery limits were clearly investigated; neither charging
costs of EV nor operation costs of DSOs was a research subject. Reference [6] aimed to minimize system operation
cost considering EV owners’ satisfaction level. In addition, the prediction function of EVs’ behavior and energy
demand is formulated as an important component. In [7], charging strategies with several commonly studied
objectives were compared. Based on simulation results, the strategies of different objectives were evaluated in
delivered energy volume and energy costs. Besides, to maximize the usage of renewable generations by charging
EVs was also investigated by [8]. This is another research subject especially in the context of micro-grids. Other
strategies from an aggregator’s point of view often share a common objective, which is to minimize the energy
procurement costs for charging EVs. As proposed in [9~11], the goal was to achieve minimum charging cost by
optimally arranging EVs’ charging power and time. Providing the ancillary service to distribution systems, e.g.
congestion management, was considered as a system constraint.
Scheduling algorithms for EVs can optimize charging decisions for the next time step solely based on present
information on the distribution system and EVs. This scheme is only suitable for objectives without the factor of
time, e.g. the grid capacity or charging demand of EVs [4, 5]. On the other hand, implementations with a global
scale over long periods are demonstrated by [7] and [10]. They emphasized that the system optimality can be
achieved under the assumption of a complete knowledge in the studied period. However, this assumption is not
realistic and therefore only possible in offline analysis. To overcome this limitation and to guarantee the online
applicability of a strategy, moving window scheduling algorithms are introduced by [6], [8],[9] and [11]. These
strategies had in common that they required forecast over a fixed time horizon. Charging decisions over the next
short period were made based on the forecasting data. During the execution of the algorithm, information and
solutions were updated in a moving window pattern.
In this study, we implement and improve the moving window optimization method for scheduling EV charging.
A binary linear optimization strategy is implemented in a centralized scheme, where distribution grid limits and
EV’s charging demand are both taken as constraints. The focus of this work is to evaluate the effect of the selection
of the optimization window, which is carried out by numerical simulations. We analyzed this effect in terms of costs
and technical variables. Lastly, impacts of the forecasting error on the optimization result are also quantitatively
analyzed.
This paper is outlined as follows. Section 2 describes the proposed charging scheduling algorithm. In Section 3,
the system model and simulation parameters are presented. Evaluations on scheduling strategies are performed based
on simulations on specific scenarios. Finally, a short conclusion and discussion is given in Section 5.

2. Problem formulation

Considering forecasting data of electricity price and driving information of EVs, we proposed a centralized
strategy with the objective to minimize the charging cost of an EV fleet consisting of multiple vehicles. In this
section, the objective function and the associated boundary conditions are first discussed. The optimization scheme
and the work flow of the scheduling algorithm are presented subsequently.
 Chenjie Ma et al. / Energy Procedia 73 (2015) 173 – 181 175

2.1. Objective function

The EV scheduling algorithm optimizes the charging set-points of an EV fleet in a discrete manner. As input
data, the proposed algorithm requires a prediction of total load in the system under investigation, a prediction of
electricity price over a fixed horizon as well as the energy demand of EVs and desired departure time estimated by
users. The main objective of the algorithm is to minimize the total charging costs of all EVs based on online energy
price forecast. We consider grid operation limits, limits of charging time window determined by EV users and
energy demand of individual EVs as constraints. The optimization problem is formulated as a binary linear
programming problem and solved using the OPTimization Interface (OPTI) Toolbox [12]. The overall objective
function can be expressed as follows:

min f ¦¦ c
dD nN
d ˜ PEV ,n ˜ Sd ,n ˜ 't  (1)

where
D set of all time steps within the optimization window;
N set of all EVs in the fleet;
cd forecast electricity price at time step d;
PEV,n rated charging power of the nth EV, assumed as constant quantities;
Sd,n charging schedule of the nth EV at time step d (0 for not charging, 1 for charging);
∆t length of a time step.
For technical constraints, we first consider the grid operation limit, which is defined in this work as the annual
maximum load (Plim) without EV. As described by Eq. (2), the sum of total charging power and the total demand of
loads at each time step must remain lower than the permissible power limit. Here, L is the set of all loads in the grid;
Pload,l stands for the power demand of the lth load.

¦ P
nN
EV ,n ˜ Sd ,n   ¦ Pload,l d d Plim , d  D
lL
(2)

The availability condition of charging an EV is formulated as Eq. (3). A schedule is valid, only if the EV is
connected to the grid at the corresponding time step.

­0, when EV is not connected to the grid

Sd , n d ® ,d  D & n N (3)
¯1, when EV is connected to the grid

Finally, Eq. (4) describes the constraints of energy demandof the EV battery. At the end of a scheduling interval,
all EVs should be fully charged. EEV,ndenotes the energy demand of an EV defined by initial SOC level at the
beginning of a scheduling window.

¦ P
dD
EV ,n ˜ Sd ,n ˜ 't  t EEV ,n , n  N (4)

2.2. Optimization scheme

In order to account for uncertainties of user behavior as well as system load and market prices, our charging
scheduling strategy is implemented as a moving window optimization scheme. In Fig. 1, the temporal behavior of
this scheme is demonstrated. Here, the optimizer requires the aforementioned forecast information of the future over
a fixed horizon, which can span over several time steps. Based on the online prediction data, the algorithm optimizes
charging set-points of EVs for this scheduling window. As the time advances, the algorithm updates the predicted
176 Chenjie Ma et al. / Energy Procedia 73 (2015) 173 – 181

information and determines charging schedules for the next window. Although schedules are made over several time
steps, only charging decisions are carried out in the present time step, which is marked in blue in the figure. This
forward looking character takes future uncertainty into consideration by making decision for the current step; its
constantly window-moving scheme reduces the demand of forecasting information. In opposite, global optimization
schemes from literature review assumed a complete knowledge about the future. Their determined schedule would
be valid for the complete time span.

6FKHGXOLQJZLQGRZ
7LPH
W W
W W W
7LPHVWHS

7LPH
W W
W W W
 
 
 
 

7LPH
W WQ
W W W WQ

Fig. 1.Scheme of moving window optimization

3. System modeling

In this work, we investigate a synthetic low voltage system with 50 % penetration of EVs. Applied data source
and models are presented in this section.

3.1. EV charging model

EV charging models are derived from our previous work [3], where detailed physical models of EV charger and
Li-Ion battery packs are established and validated. In this work, EV models are also parameterized based on the
specifications of Mitsubishi i-MiEV[13]. It is powered by a 16 kWh Li-ion battery pack. For charging device, a
standard power rate of 3.7 kW is assumed.

3.2. EV behavior model

The user behavior of EVs in this work is based on a surveyed dataset, which is collected by the project “Mobility
in Germany 2008” [14]. This survey covered driving behavior of thousands of households over one year. Exact trips
on workdays and weekends and the energy consumption are recorded in this database. In the simulation, trips are
randomly selected from the database and assigned as driving behavior for individual EVs.

3.3. Power consumption and energy price

A winter week (third week of 2013) is selected as the study period. A set of household load profiles is generated
based on German statistics [15]. These profiles describe the electric power consumption of 146households of
different types with time resolution of 1 minute. We use the 1 h average value of intraday price profile from the
 Chenjie Ma et al. / Energy Procedia 73 (2015) 173 – 181 177

EEX spot market [16]. The price profiles in a 24 h period of a workday and a weekend (Saturday) scenario are
shown in Fig. 2.

100 55
Energy price [€/MWh]

Energy price [€/MWh]

50
80

45
60
40

40
35

20 30
10 12 14 16 18 20 22 24 2 4 6 8 10 10 12 14 16 18 20 22 24 2 4 6 8 10
Time [h] Time [h]
a) Workday b) Weekend
Fig. 2. Energy profile in a 24 h period

4. Results of the fleet optimization

The proposed scheduling algorithm is tested on the generic grid model with a high EV penetration ratio of 50 %.
Typical workday and weekend scenarios are simulated with different optimization parameters. In this section,
impacts of schedule window size and forecasting error on charging costs are discussed.

4.1. Coordinated and uncoordinated charging

Figure 3 illustrates the charging behavior of EVs and total power demand in a 24 hour period of a typical
workday and a weekend day scenario. On the workday, the uncoordinated charging scheme indicates charging of
EVs immediately after they are connected to the grid. This yields severe peaks of total power demand between 6 and
8 pm. The coordinated charging strategy schedules the charging of EVs during midnight as it is sensitive to the
energy prices. On the weekend, EVs stay plugged in longer and more often during daytime than in a workday. This
provides a higher flexibility for the scheduling algorithm. Accordingly, the achieved solution avoids charging EVs
during peak hours of high demand and shifts the charging to hours of low energy prices (between 11 am and 9 pm).

4.2. Impact of forecasting error on moving window optimization results

A sensitivity analysis of the optimization results on imperfect price forecasts is performed in this part. It is assumed
that the forecasting error subjects to a normal distribution with increasing standard deviation as a function of time.
We exemplarily study hourly increments of 0.5 % and 1 % on forecasting error for the price profile. This
corresponds to error bands of 12 % and 24 % for a forecast in the coming 24th hour. The optimal charging costs
calculated for 50 simulations are presented in Fig. 4. In the workday scenario, we observe that the lowest charging
costs are reached at a larger scheduling window if the price forecast contains errors, as opposed to simulations with
a perfect forecast. The error adds more uncertainty at a larger window size, so that the obtained optimal charging
costs increase if the window size further increases. However, this effect is, although still visible, less significant on
the weekend day. In addition, the global optimization scheme is more sensitive to error, as it achieves much higher
costs for both of the scenarios. The difference between two error bands is marginal.
178 Chenjie Ma et al. / Energy Procedia 73 (2015) 173 – 181

500 500
Accumulated power
Power demand [kW] demand of charging Power demand of

Power demand [kW]

400 400 charging individual EV
in different color

300 300

Load profile
200 200

100 100

0 0
10 12 14 16 18 20 22 24 2 4 6 8 10 10 12 14 16 18 20 22 24 2 4 6 8 10
Time [h] Time [h]
70 70
Charging schedule
60 60
Available for charging
50 50
EV index

EV index
40 40
30 30
20 20
10 10

10 12 14 16 18 20 22 24 2 4 6 8 10 10 12 14 16 18 20 22 24 2 4 6 8 10
Time [h] Time [h]
a) Workday, uncoordinated b) Workday, coordinated
500 500
Power demand of
Power demand [kW]

Power demand [kW]

400 Accumulated power

400 charging individual EV
in different color
demand of charging
300 300
Load profile
200 200

100 100

0 0
10 12 14 16 18 20 22 24 2 4 6 8 10 10 12 14 16 18 20 22 24 2 4 6 8 10
Time [h] Time [h]

70 70
Charging schedule
60 Available for charging 60
50 50
EV index

EV index

40 40
30 30
20 20
10 10

10 12 14 16 18 20 22 24 2 4 6 8 10 10 12 14 16 18 20 22 24 2 4 6 8 10
Time [h] Time [h]
c) Weekend, uncoordinated d) Weekend, coordinated
Fig. 3.Uncoordinated and coordinated charging on workday and weekend scenarios
 Chenjie Ma et al. / Energy Procedia 73 (2015) 173 – 181 179

43
Error increment: 0.5%/h
Error increment: 1%/h

Total charging cost, [€]

42

41

40

39

38 Global
7 9 11 13 15 17 19 21 23
Optimization window, [h] optimization

b) Weekend
49
Error increment: 0.5%/h
48 Error increment: 1%/h
Total charging cost, [€]

25-75th
47 percentile
Median
46

45

44

43 Global
7 9 11 13 15 17 19 21 23
optimization
Optimization window, [h]
a) Workday
Fig. 4.Optimal charging costs with imperfect forecasting data

4.3. Impact of scheduling window size

In order to analyze the impact of selecting the moving window, simulations on different window size from 1 to
24 hours are carried out with perfect forecasting data. Comparisons of scheduling window sizes against the global
optima are presented in Fig. 5(the global optimum is the optimum solution obtained by a global optimization scheme
mentioned in Section 2.2.) As the window size increases, the optimal charging costs obtained by the moving
window scheme approach to the global optima obtained by the global optimization scheme on both workday and
weekend scenarios. Although different convergence characteristics are observed, the moving window scheme is able
to provide solutions as good as the global optimization scheme. On the specific test scenarios, optimal solutions are
achieved by a window size of 12 h and 19 h, respectively. In practice, EVs availability range and the shape of price
profiles should be considered by optimal selection of the window size.
180 Chenjie Ma et al. / Energy Procedia 73 (2015) 173 – 181

60 50

55
Total charging cost, [€]

Total charging cost, [€]

Uncoordinated charging
45
50 Uncoordinated charging

45 40

40
Global optimum Global optimum
35
35

30 30
7 9 11 13 15 17 19 21 23 7 9 11 13 15 17 19 21 23
Optimization window, [h] Optimization window, [h]
a) Workday b) Weekend
Fig. 5.Comparison of optimal charging costs of one day at different scheduling windows
(The solver is incapable to obtain viable solutions with a window size fewer than 7 h.)

5. Conclusion

This work proposed an implementation of scheduling strategy based on moving window optimization scheme.
This strategy shows generally a fast convergence characteristic and is more robust against the global optimization
scheme. These advantages enable a reliable determination of optimum charging schedules in low costs. Also,
moving window optimization scheme is suitable for online applications due to continuously information update
pattern and a fixed forecasting horizon. For a real application, a window size between 12 and 19 h can be
recommended from test results. If forecasting data contain errors, the selection of optimal window size requires
more detailed studies. Based on the exemplary test cases, a window size of 17 h shows lowest charging costs in both
workday and weekend scenarios. To consolidate the implemented algorithms, more simulation tests on real
scenarios with longer duration will be studied in the next steps. Technical impacts on distribution grids will also
studied using detailed system simulation models. Finally, the real time performance of the strategy requires more
attention.

Acknowledgements

The presented paper is based on research performed in the project ‘Systemintegration von Elektromobilität
(SIEM)’ (FKZ 0325402). The authors thank the German Federal Ministry for Economic Affairs and Energy
(BMWi) and the ProjektträgerJülich (PTJ) for supporting this project and take full and sole responsibility for the
content of this paper.

References

[1]S. Shafiee, M. Fotuhi-Firuzabad, and M. Rastegar, “Investigating the impacts of plug-in hybrid electric vehicles on power distribution
systems,” Smart Grid, IEEE Transactions on, vol. 4, no. 3, pp. 1351–1360, Sept 2013.
[2] A. Masoum, S. Deilami, P. Moses, M. Masoum, and A. Abu-Siada, “Smart load management of plug-in electric vehicles in distribution and
residential networks with charging stations for peak shaving and loss minimisation considering voltage regulation,” Generation, Transmission
Distribution, IET, vol. 5, no. 8, pp. 877–888, August 2011.
[3] C. Ma, F. Marten, J.-C. Töbermann and M. Braun, “Evaluation of modeling and simulation complexity on studying the impacts of electrical
vehicles fleets in distribution systems,” 18th Power Systems Computation Conference (PSCC), Wroclaw, Poland, August 18-22, 2014.
[4] P. Richardson, D. Flynn, and A. Keane, “Optimal charging of electric vehicles in low-voltage distribution systems,” Power Systems, IEEE
Transactions on, vol. 27, no. 1, pp. 268–279, Feb 2012.
 Chenjie Ma et al. / Energy Procedia 73 (2015) 173 – 181 181

[5] K. Zhou and L. Cai, “Randomized PHEV charging under distribution grid constraints,” Smart Grid, IEEE Transactions on, vol. 5, no. 2, pp.
879–887, March 2014.
[6] M. Shaaban, M. Ismail, E. El-Saadany, and W. Zhuang, “Real-time PEV charging/discharging coordination in smart distribution systems,”
Smart Grid, IEEE Transactions on, vol. 5, no. 4, pp. 1797–1807, July 2014.
[7] I. Sharma, C. Canizares, and K. Bhattacharya, “Smart charging of PEVs penetrating into residential distribution systems,” Smart Grid, IEEE
Transactions on, vol. 5, no. 3, pp. 1196–1209, May 2014.
[8] W. Leterme, F. Ruelens, B. Claessens, and R. Belmans, “A flexible stochastic optimization method for wind power balancing with PHEVs,”
Smart Grid, IEEE Transactions on, vol. 5, no. 3, pp. 1238–1245, May 2014.
[9] A. O’Connell, D. Flynn, and A. Keane, “Rolling multi-period optimization to control electric vehicle charging in distribution networks,”
Power Systems, IEEE Transactions on, vol. 29, no. 1, pp. 340–348, Jan 2014.
[10]P. Rezaei, J. Frolik, and P. Hines, “Packetized plug-in electric vehicle charge management,” Smart Grid, IEEE Transactions on, vol. 5, no. 2,
pp. 642–650, March 2014.
[11]Y. He, B. Venkatesh, and L. Guan, “Optimal scheduling for charging and discharging of electric vehicles,” Smart Grid, IEEE Transactions
on, vol. 3, no. 3, pp. 1095–1105, Sept 2012.
[12]J. Currie and D. I. Wilson, "OPTI: Lowering the Barrier between Open Source Optimizers and the Industrial MATLAB User," Foundations
of Computer-Aided Process Operations, Georgia, USA, 2012.
[13] Mitsubishi Motors, Dec 2014, http://www.mitsubishicars.com/imiev.
[14]INFAS, DLR, and DIW, “Mobility in Germany (MiD) 2008”, 2008.
[15]J. von Appen, T. Stetz, M. Braun, and A. Schmiegel, “Local voltage control strategies for PV storage systems in distribution grids,” Smart
Grid, IEEE Transactions on, vol. 5, no. 2, pp. 1002–1009, March 2014.
[16]European Energy Exchange (EEX), https://infoproducts.eex.de/.