Vous êtes sur la page 1sur 11

156 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO.

1, JANUARY 2018

A Novel Markov-Based Temporal-SoC Analysis


for Characterizing PEV Charging Demand
Siyang Sun, Qiang Yang , Member, IEEE, and Wenjun Yan

Abstract—The integration of a massive number of plug-in current power distribution networks brings direct challenges for
electric vehicles (PEVs) into current power distribution net- network planning, control, and management [4]. The temporal
works brings direct challenges to network planning, con- and spatial distribution characteristics of PEV charging demand
trol, and operation. To increase the PEV penetration level
with minimal negative impact, the dynamical PEV travel imposes a significant impact on the load profiles of distribution
behaviors and charging demand need to be better under- networks, and hence results in unexpected voltage fluctuation
stood. This paper presents a Markov-based analytical ap- and increased peak-valley gap. In the presence of the increasing
proach for modeling PEV travel behaviors and charging penetration level of PEVs, the PEV travel behaviors and charg-
demand. The travel behaviors of individual PEVs are ex-
ing demand need to be better understood and explicitly formu-
pressed mathematically through Monte Carlo simulation
considering two essential factors: temporal travel purposes lated to ensure the reliable and economical network operation.
and state of charge (SoC). Markov model and hidden Markov However, due to the diverse PEV travel patterns and charging
model (HMM) are adopted to explicitly formulate the prob- requirements, the charging behaviors of individual PEVs are
abilistic correlation between multiple PEV states and SoC stochastic and hard to be predicted, making the modeling of
ranges. This modeling approach provides an efficient and
PEV charging demand a nontrivial task [5].
generic tool for analyzing PEV travel behaviors and charg-
ing demand based on available PEV statistics. The analyt- In the literature, a collection of research effort has been made
ical model is further adopted in the impact assessment of to address the challenges of modeling PEV charging demand,
two PEV normal charging scheduling strategies for a range attempting to formulate the charging behaviors, e.g., charging
of PEV penetration levels in an IEEE 53-bus test network start time, charging duration, and initial state of charge (SoC)
with field data (network parameters and realistic PEV statis-
of charging (e.g., [6]–[12] and [18]). The charging start time
tics). The results demonstrate the benefit of the proposed
modeling approach in network analysis considering PEV was considered to be daily home arrival time in [6]–[9] based
integration. on the survey statistics, or assumed to follow a given normal
distribution in [10]. Daily driving distance was generally used
Index Terms—Charging demand, distribution networks,
hidden Markov model (HMM), Markov model, Monte Carlo to calculate the energy depletion of battery, which represents
simulation, plug-in electric vehicle (PEV), state of charge the charging duration in [6]–[11], or the initial SoC of charging
(SoC). was assumed to be known as a priori from historical statistics
in [12]. For the charging behaviors, the solutions in [7] and
I. INTRODUCTION [8] assumed that PEVs did not need to be charged every
HE growing concerns of fossil fuel consumption and day, and the study in [12] considered that the PEVs are not
T greenhouse gas emission have motivated the quick devel-
opment of plug-in electric vehicles (PEVs). It is anticipated that
necessarily to be fully charged. The probabilistic power flow
analysis was applied to model PEV charging demand of in
PEV will be a fairly large segment of US national fleet (at least [11], and a spatial-temporal model was proposed based on
10% by 2020 and 50% by 2050 of new car sales projection) origin-destination analysis in [18].
[1]. Currently, PEVs can be charged either in normal charg- The Markov model and hidden Markov model (HMM) are
ing mode at parking lots, or in fast charging mode at charging well known to describe transition process among multiple dis-
stations with the typical charging duration of 20–30 min [2]. crete states. A number of studies have adopted Markov theory
However, the integration of a massive number of PEVs into for characterizing PEV charging demand (e.g., [13]–[15]). In
[13], Markov chain was used to evaluate the impact of PEV
charging on residential power demand profile. Sundström et al.
Manuscript received February 10, 2017; revised May 24, 2017; ac- [14] proposed a semi-Markov model to predict the PEV travel
cepted June 20, 2017. Date of publication June 28, 2017; date of
current version January 3, 2018. This work was supported by the behaviors, e.g., the next arrival location and the waiting time at
Zhejiang Provincial Natural Science Foundation of China under Grant the current location. The PEV charging events were modeled as a
Z15E070001. Paper no. TII-17-0251. (Corresponding author: Qiang Markov decision process with unknown transition probabilities
Yang.)
The authors are with the College of Electrical Engineering, Zhe- to appropriately schedule the charging behaviors of individual
jiang University, Hangzhou 310027, China (e-mail: sunsiyang911@zju. PEVs in [15].
edu.cn; qyang@zju.edu.cn; yanwenjun@zju.edu.cn). However, these aforementioned PEV charging demand mod-
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. els have some obvious limitations. First, most of the solutions
Digital Object Identifier 10.1109/TII.2017.2720694 considered that PEVs are charged only at home after daily trips.

1551-3203 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
SUN et al.: NOVEL MARKOV-BASED TEMPORAL-SOC ANALYSIS FOR CHARACTERIZING PEV CHARGING DEMAND 157

However, the daily driving distance may be more than PEV TABLE I
PEV CLASSIFICATION AND PARAMETERS
maximum mileage or the SoC at departure time is not full, re-
quiring the PEV to be charged immediately at a fast charging
station. Thus, the fast charging mode should be taken into con- PEV type M1 N1
sideration in modeling of PEV demand. Furthermore, most of Battery capacity μM 1 = 28.5 σ M 1 = 14.7 μN 1 = 23.0 σ N 1 = 9.5
C bat C bat Cbat C bat
the existing solutions assumed that the PEV charging demand C bat (kWh) M1
Cm = 72.0 M1
Cm = 10.0 C m
N 1 = 40.0 C N 1 = 9.6
ax in ax m in
merely depends on the home arrival time and daily driving dis- Energy E cM 1 = 0.159 E cN 1 = 0.185
tance, and the temporal characteristics of PEV travel (e.g., de- consumption per
kilometer
parture time, arrival time), which can directly affect the temporal E c (kWh/km)
characteristics of PEV travel behaviors and charging demand, Market share ms msM 1 = 89% msN 1 = 11%
is ignored for sake of simplicity. In addition, there is still a lack
of models to explicitly formulate the PEV travel behaviors and
transition dynamics of operational states. To the author’s best by companies or unemployed people, i.e., other purpose PEVs
knowledge, the PEV charging demand model that considers the (O-PEVs), which may have multiple trips in a day.
correlation between PEV states and SoC has not been reported Two PEV types (M1, N1) are considered according to a survey
yet. of the European PEV market [19]. The battery capacity Cbat of
To this end, this paper addressed the challenges of charging PEVs is assumed to follow the normal distribution N (μ, σ 2 ),
demand modeling and presented a novel Markov-based analyt- as given in (1). In this paper, individual PEVs are considered
ical approach. The main contributions made in this paper are for one of the purposes (C-PEV or O-PEV) and types (M1 or
as follows. First, the operational travel behaviors of individ- N1), and hence the travel patterns and PEV specifications can
ual PEVs are mathematically formulated through Monte Carlo be characterized. The key PEV parameters and statistics are
simulation based on the temporal-SoC analysis to obtain the provided in Table I [18], [19]:
expected probabilities of multiple PEV states over a day. The
1 ( x −μ ) 2
transition probability matrix of Markov model and the confu- f (x) = √ e− 2σ 2 . (1)
sion matrix of HMM are derived to describe the probabilistic 2πσ
correlation among multiple PEV states and SoC ranges in any
given time slots. Second, the proposed model is further adopted
and validated with field statistics in assessing the impact of two B. PEV Charging Power Level
PEV normal charging scheduling strategies through simulation Currently, two PEV charging modes are generally adopted:
experiments carried out for an IEEE 53-bus test network. It normal charging mode and fast charging mode with the typical
shows that the “smart” charging strategy can effectively imple- charging power Pn − ch and Pf − ch of 3.3 kW (220 V/15 A)
ment peak shaving and valley filling via scheduling the PEV and 50 kW (400 V/125 A), respectively [11].
charging behaviors. In fact, the proposed model is considered
a generic analytical tool for PEV charging demand characteri- C. PEV Temporal Travel Patterns
zation. It can be adopted for a variety of network management
tasks considering PEV penetration, e.g., energy dispatch, charg- Based on the available statistics of typical PEV travel times,
ing station planning. such as National Household Travel Survey (NHTS) [3], [4],
The remainder of this paper is organized as follows: [17], [20], the departure and arrival times of C-PEVs in the
Section II presents the PEV statistics and a set of assumptions; morning (tm d and tm a ) and evening (ted and tea ) can be obtained
Section III presents the proposed Markov-based modeling ap- through data normalization, maximum likelihood estimation,
proach for PEV charging demand through the temporal-SoC and curve-fitting. In this paper, the departure and arrival times
analysis, with the mathematical validation; Section IV applies (tm d ,tm a , ted , and tea ) of C-PEVs can be modeled following the
the proposed modeling approach to evaluate the impact of PEV normal distribution (1) with the mean time μ and the standard
normal charging demand on IEEE 53-bus test network; finally deviation σ of (6:52, 8:00, 16:52, 17:29) and (1.3, 3.4, 2.3, 3.25),
a set of conclusive remarks are given in Section V. respectively. Similarly, the departure times td of O-PEVs also
follow the normal distribution (μ = 13 : 51 and σ = 5.2), the
arrival times ta are as same as those of C-PEVs.
II. PEV CHARACTERISTICS AND ASSUMPTIONS
The PEV specifications and a set of assumptions in this paper D. Assumptions
are first discussed as follows:
To analyze and characterize the PEV charging demand, the
following assumptions are made.
A. PEV Classification 1) The PEV travel patterns are as same as the conventional
Currently PEVs can be characterized and classified from dif- internal combustion engine (ICE) vehicles.
ferent aspects. It is reported in [18] that 61% (rC ) of vehicles 2) For O-PEVs, they are assumed to drive during the depar-
in the UK are primarily used for commuting between home and ture time td and the arrival time ta .
workplaces, i.e., commuting PEVs (C-PEVs), which have two 3) The operational PEV states are divided into four dif-
trips in a day; and the other vehicles (rO = 39%) are owned ferent kinds based on the characteristics of PEV travel
158 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 1, JANUARY 2018

behaviors, i.e., normal charging state (Sn − ch ), fast


charging state (Sf − ch ), driving state (Sd ), and parking
state (Sp ), respectively.
4) For the sake of battery lifetime protection, the PEVs are
not allowed to be charged (or discharged) once the SoC
reaches 0.8 (or 0.2); PEV starts to be charged immediately
upon the arrivals at home in normal mode, unless the SoC
reaches 0.8 or it needs to leave home.
5) PEVs are charged with the constant power rate (as
adopted in [16]) and the operational state remains un-
changed within individual time slots (30 min per time
slot, i.e., in total 48 time slots in a day).

III. MARKOV-BASED TEMPORAL-SOC ANALYSIS OF PEVS


The temporal travel behavior and SoC can directly affect its
charging demand. This paper adopts an analytical temporal-SoC
analysis for PEV travel and charging demand characterization,
including three key steps: Monte Carlo simulation of PEV travel
behaviors; calculation of Markov model and HMM; and deter-
mination of the expected probability distribution of PEV states,
as illustrated in Fig. 1.

A. Monte Carlo Simulation of PEV Travel Behaviors


1) PEV parameters: the Monte Carlo simulation is used to
generate the battery capacity Cbat of individual PEV sam-
ples based on the probability distribution and constraints
(see Table I). In the case that the generated capacity is
not within the maximum (Cm ax ) and minimum (Cm in )
kWh constraints, the process will be repeated until the
constraints are satisfied.
2) PEV purposes of usage: the generated PEV samples are
determined to be C-PEV or O-PEV according to the rate
statistics (rC and rO ) in [18].
3) PEV travel times: Nr random samples of PEV travel times
according to the distribution statistics are generated, re-
spectively, through Monte Carlo simulation, from which Fig. 1. Markov-based modeling approach for temporal PEV states.
Nd valid time samples (e.g., the arrival time should be
later than the departure time) can be further extracted.
4) PEV travel behavior simulation of individual PEV sam- of time samples. More specifically, there are two cases in this
ples: two steps are required in the proposed temporal-SoC scenario based on the SoC.
analysis to determine the travel behaviors of PEV sam- Case 1.1: [SoC(t) > 0.2]: PEV is in driving state (Sd ) and
ples in each time slots. Step 1 considers the current time its SoC decreases with a certain rate, which is determined by
slot to preliminarily examine the temporal travel purpose the battery capacity (Cbat ), energy consumption per kilometer
of individual PEV samples (i.e., scenario 1-6 for C-PEV (Ec ), and the average driving speed (v = 46.45 km/h obtained
and scenario 1-4 for O-PEV); and step 2 further deter- from NHTS [3]). Thus, the initial SoC of the next time slot
mines the PEV travel behavior based on the current SoC SoC(t + 1) can be expressed as (2), where SoC(t) is the initial
according to the assumption (4) given in Section II-D SoC of the current time slot t. As the time slot period is half an
(e.g., case 1.1 or 1.2). The detailed modeling process for hour, there will be a “2” in the following equations:
determining the travel behavior of C-PEV and O-PEV is vEc
described as follows. SoC(t + 1) = SoC(t) − . (2)
2Cbat
A. 1) For Any C-PEV: According to the temporal travel
purposes, there are in total six different temporal scenarios. Case 1.2: [SoC(t) ≤ 0.2]: the PEV needs to be charged at
Scenario 1 [t < tea (j − 1) − 24)]: the evening arrival time a fast charging station immediately so that the state switches
of the last day is later than 24:00, which means the PEV has not to fast charging (Sf − ch ). The initial SoC in the next time slot
arrived at home and still needs to drive, where j is the number SoC(t + 1) is given in (3), where Pf − ch = 50 kW is the rated
SUN et al.: NOVEL MARKOV-BASED TEMPORAL-SOC ANALYSIS FOR CHARACTERIZING PEV CHARGING DEMAND 159

power of fast charging:


Pf − ch
SoC(t + 1) = SoC(t) + . (3)
2Cbat
Scenario 2 [tea (j − 1) − 24 < t < tm d (j)]: the PEV has al-
ready arrived at home in the previous evening, and has not left
to work in the morning yet. There are also two cases in this
scenario based on the SoC in detail.
Case 2.1: [SoC(t) < 0.8]: the PEV state switches to the nor-
mal charging state (Sn − ch ). The initial SoC of the next time
slot SoC(t + 1) can be expressed as (4), where Pn − ch = 3.3
kW is the rated power of normal charging:
Pn − ch
SoC(t + 1) = SoC(t) + . (4)
2Cbat
Case 2.2: [SoC(t) ≥ 0.8]: the PEV parks at home with the
state of parking (Sp ), and the SoC remains unchanged

SoC(t + 1) = SoC(t). (5)

Scenario 3 [tm d (j) < t < tm a (j)]: the PEV has departed Fig. 2. Illustration of different temporal scenarios for C-PEV and
from home to work in the morning and has not arrived at O-PEV.
the workplace. In this scenario, the two cases are as same as
Scenario 1, based on the SoC of the PEV at the current simula-
PEV travel behavior using (2)– (5). Such process is repeated for
tion time slot t.
individual time slots over a day.
Scenario 4 [tm a (j) < t < ted (j)]: the PEV has arrived at the
Here, it is considered that the initial SoC at the start of ev-
workplace and has not departed to home in the evening. Here,
ery travel time samples (0:00) is equal to the SoC value at the
the PEV parks at the workplace, i.e., (Sp ) and the SoC remains
end of the previous time sample (24:00), i.e., all generated time
unchanged, as shown in (5).
samples are within the same day so that the algorithm can con-
Scenario 5 [ted (j) < t < tea (j)]: the PEV has departed from
tinuously simulate the PEV battery consumption throughout the
the workplace in the evening and has not arrived at home. In this
overall time samples. Thus, the modeling approach can simulate
scenario, the two cases are also as same as Scenario 1, based on
for a PEV over Nd days successively and the initialization of
the SoC of the PEV at the current simulation time slot t.
SoC only needs to be carried out once (the initial SoC is assumed
Scenario 6 [tea (j) < t < tm d (j + 1) + 24]: the PEV has al-
to be 0.5) for individual PEV samples. Based on Monte Carlo
ready arrived at home in the evening and has not left to work
simulation technique, the aforementioned steps are repeated for
in the next morning. This scenario has two cases, which are as
each PEV sample in each time slot until the termination condi-
same as scenario 2.
tions are met.
The aforementioned temporal scenarios of C-PEVs are illus-
trated in Figs. 2 and 3(a), where Sn − ch ,Sf − ch ,Sd , and Sp rep-
resent four PEV states, i.e., normal charging state, fast charging B. Markov-Based Model of PEV
state, driving state, and parking state, respectively, and 2)–(5) in To formulate the PEV travel behaviors and charging demand,
Fig. 3(a) represents equations (2)–(5), respectively. the average transition probabilities among multiple PEV states
A. 2) For Any O-PEV: There is no arrival time in the morn- as well as the probabilistic correlation between PEV states and
ing tm a and no departure time in the evening ted . The temporal SoC in individual time slots needs to be studied. As the PEV
scenarios for O-PEVs can be described by certain scenarios of state transition probability is independent from the initial state
C-PEVs, as illustrated in Figs. 2 and 3(b). and the previous transition process, i.e., a Markov process, the
Scenario 1 [t < ta (j − 1) − 24]: it is as same as Scenario 1 Markov model is adopted to describe the average transition
of C-PEV. probabilities among multiple PEV states in any given time slot.
Scenario 2 [ta (j − 1) − 24 < t < td (j)]: it is equivalent to The transition process among four PEV operational states and
scenario 2 of C-PEV. the Markov transition probability matrix P are illustrated and
Scenario 3 [td (j) < t < ta (j)]: it is equivalent to scenario 1 given in Fig. 4 and (6), respectively,
and 5 of C-PEV. ⎡ ⎤
Scenario 4 [ta (j) < t < td (j + 1) + 24]: it is equivalent to p11 p12 p13 p14
scenario 6 of C-PEV. ⎢ ⎥
⎢ p21 p22 p23 p24 ⎥
Based on the previous steps, once the travel behavior of indi- P =⎢ ⎢p

⎥ (6)
vidual PEV samples in a time slot is determined, the SoC of the ⎣ 31 p32 p33 p34 ⎦
next time slot can be predicted and updated based on the current p41 p42 p43 p44
160 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 1, JANUARY 2018

Fig. 4. Markov model of PEV.

Fig. 5. Markov transition probability matrix of PEV.

However, some state transition process may not occur among


some certain PEV states in some certain time slots in the simula-
tion, which needs to be removed from the calculation of Markov
transition probability matrix.
Through the Markov transition probability matrix, the average
transition probabilities among multiple PEV states in individual
time slots can be explicitly determined. Thus, given an initial
Fig. 3. Flowchart of PEV travel behavior modeling (a) C-PEV; PEV state distribution, the expected distribution of PEV states
(b) O-PEV.
in any time slots can be predicted. On the other hand, HMM
is a special Markov model in which the system being modeled
where pij represents the transition probability from state i to is assumed to be a Markov process with unobserved (hidden)
state j. As there are in total 48 time slots per day, each element states. In this paper, HMM is adopted to describe the proba-
of the Markov transition probability matrix is a 1 × 48 matrix. bilistic correlation between states and SoC of PEV. Here, three
The derived Markov transition probability matrix P of PEVs different SoC ranges (i.e., x0 - - x1 , x1 - - x2 , and x2 - - x3 ) are
is shown in Fig. 5. It should be noted that the obtained Markov considered, and the HMM of PEVs is illustrated in Fig. 6. In
transition probability matrix from Monte Carlo simulation rep- fact, more SoC ranges can be adopted in accordance with the
resents the average transition probabilities among PEV states. required accuracy of HMM.
SUN et al.: NOVEL MARKOV-BASED TEMPORAL-SOC ANALYSIS FOR CHARACTERIZING PEV CHARGING DEMAND 161

Fig. 6. Illustration of HMM of PEV.

The confusion matrix C of HMM is used to represent the


probabilistic correlation between the observable states (PEV
states) and the hidden states (SoC ranges), as given in (7). In
the confusion matrix, cij represents the probabilistic correlation Fig. 7. Confusion matrix of HMM.
between the observable state i and the hidden state j, where the
elements in rows and columns of the confusion matrix represent
the observable states (PEV states) and the hidden states (SoC
ranges), respectively. As there are in total 48 time slots (30 min
per time slot) over a day, each element of the confusion matrix
is a 1 × 48 matrix, too:
⎡ ⎤
c11 c12 c13
⎢ c21 c22 c23 ⎥
C=⎢ ⎣ c31 c32 c33 ⎦ .
⎥ (7)
c41 c42 c43
The confusion matrix C can also be obtained based on the
analytical approach presented in Section III (see Fig. 1). Fig. 7 Fig. 8. Probability curves of different PEV states over a day.
presents the confusion matrix of HMM over a day for multiple
PEV states for three different SoC ranges (i.e., 0.2–0.4, 0.4–0.6,
and 0.6–0.8), respectively. It should be noted that, for the curves SoC ranges (SoCt10 ,SoCt20 ,SoCt30 ) at t0 can be obtained based on
of Sf − ch shown in Fig. 7, the probabilities for two cases, (SoC: the HMM of PEV as
0.4–0.6) and (SoC > 0.6), are zero and the probability under the (SoCt10 , SoCt20 , SoCt30 ) = (Snt 0− c h , Sft 0− c h , Sdt 0 , Spt 0 ) × C (8)
case of (SoC < 0.4) is always equal to 1 over the day. This is
reasonable as once a PEV is in the fast charging mode, its SoC where SoCt10 SoCt20 SoCt30 represents the probability distribu-
is definitely lower than 0.4, as assumed in Section II. tion of three SoC ranges (0.2–0.4), (0.4–0.6), and (0.6–0.8),
From the HMM of PEV, once the distribution of PEV states respectively.
in any time slot is given, the expected distribution of SoC can Also, the PEV states distribution in the next time slot t1
be estimated based on the confusion matrix of HMM. This can (Snt 1− c h , Sft 1− c h , Sdt 1 , Spt 1 ) can be further derived using the PEV
be useful for explicit prediction of PEV charging demand in Markov model as follows:
the following time slots, and hence optimal dispatch of PEV
charging demand or planning of PEV charging stations. (Snt 1− c h , Sft 1− c h , Sdt 1 , Spt 1 ) = (Snt 0− c h , Sft 0− c h , Sdt 0 , Spt 0 ) × P.
(9)
Through the iteration of the above process for all time slots
C. Estimation of PEV States and Charging Demand
(i.e., t0 –t48 ), the expected PEV states as well as the expected
Based on the proposed modeling approach, given the initial distribution of SoC over a day can be obtained. Fig. 8 presents
distribution of PEV states (Snt 0− c h , Sft 0− c h , Sdt 0 , Spt 0 ) at t0 , the the expected probabilities of four PEV states over a day assum-
distribution of PEV states and SoC over a day can be obtained ing that the initial distribution of PEV states at t0 = 0 : 00 is
from the Markov model and HMM of PEV. The distribution of (0.3309, 0.0084, 0.0391, 0.6218), which is also obtained from
162 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 1, JANUARY 2018

the simulation. Thus, the PEV charging demand in individual


time slots can be further derived based on the expected PEV
charging state probabilities and the rated charging power.

D. Convergence Assessment of the Proposed Algorithm


First, a series of PEV travel time samples are generated in
the simulation, from which the Nd valid samples are extracted
to represent the PEV travel pattern using the available statistics.
Based on the mathematical justifications of termination criteria
for Monte Carlo simulation (e.g., [23]), an estimation of the
mean value X̄ of an arbitrary quantity X(t) apart from a splitting
of the integral into N time steps T is given by
 T sim Fig. 9. Average PEV state probabilities with different number of
1
N
1 iterations.
X̄ = X(t)dt = Xi (10)
Tsim 0 N can well represent the characteristics of the adopted PEV travel
i=1

where Tsim is the amount of simulated time, and pattern statistics.


The proposed algorithm terminates under either of the two
Tsim = N T, N >> 1 (11) conditions: 1) the maximum number of iterations Nm ax is
 reached; and 2) the mismatch of the PEV state probability dis-
1 iT
Xi = X(t)dt. (12) tribution in individual time slots between any two sequential
T (i−1)T iterations is sufficiently small, as
 N sim −1 
The estimator for the standard deviation of the estimated mean  N s im L
σ̃ 2 is given by  i=1 i Li 
max  − i=1 <ε (16)

 Nsim Nsim − 1 
1 2
N
1
σ̃ =
2
Xi − X̄ . 2
(13) where Nsim denotes the number of iterations that have already
N − 1 N
i=1 been simulated, Li is the matrix which stores the results of the
With the standard deviation of the estimated mean value, the average PEV state probability distribution after Nsim iterations,
convergence of the Monte Carlo simulation can be calculated. and ε is the convergence factor [18]. In this paper, ε is set to be
The distribution function of the estimated mean approaches a 0.0001 and Nm ax is preliminarily set to be 10 000.
Gaussian distribution and follows the central limit theorem. The condition 1) guarantees that all of the obtained results can
Thus, the probability of the estimated mean within a confidence be converged within ±ε. Fig. 9 presents the result of the average
interval of [η − δ, η + δ] can be approximated as probabilities for four PEV states after Nsim iterations, where the
 η +δ   curves in each subfigures represent the PEV state probabilities
1 (x − η)2 for all 48 time slots. It can be observed that the results of the pro-
P η − δ ≤ X̄ ≤ η + δ = √ exp − dx
η −δ 2πσ̃ 2 2σ̃ 2 posed algorithm are converged with the increase of the number
(14) of iterations Nsim . It is also found that the termination condition
where η is the expected value of the random variable. 2) can be met when the number of iterations Nsim reaches about
Here, if the confidence interval δ is chosen to be two times 1800. Thus, the maximum number of iterations Nm ax is set to
of the standard deviation of the estimated mean (i.e., δ = 2σ̃), 20 00, in this paper, to ensure the simulation convergence with
the integral in (14) can be solved and the confidence level is acceptable computation time.
95.45%. The estimated relative error is given by There are some other sampling methods, such as Latin hy-
percube sampling and heuristic moment matching, available
δ 2σ̃ 2
rerr = = . (15) and adopted in the literature (e.g., [4] and [22]). These sam-
η X̄ pling methods are used to achieve an acceptable computational
The Monte Carlo simulation terminates in the case that the complexity than random sampling methods (e.g., Monte Carlo
0
given relative error rerr is achieved, indicating that the sampling simulation). The computational complexity of the proposed al-
results from the Monte Carlo simulation are converged within gorithm with the aforementioned parameters are evaluated using
the error range of ±rerr0
with the confidence level of 95.45%. MATLAB R2014a on a laptop computer with a 2.50 GHz In-
In this paper, Nr is set to be 20 000, and the relative errors of tel(R) Core(TM) i7-6500U CPU and a 4.00 GB RAM. The
Nd valid PEV travel time samples are calculated to be 0.0051, computation time is 445.276 s (about 7.5 min), which is consid-
0.0061, 0.0033, 0.0034, and 0.0093 for tm d ,tm a ,ted ,tea , and ered acceptable. In addition, it should be noted that the proposed
td , respectively. All of the Nd valid travel time samples are modeling approach is a generic analytical tool for PEV travel
converged within ±rerr 0
= 0.01 with 95.45% confidence interval, behavior prediction and charging demand estimation based on
which means that these Nd valid travel time samples are nearly available PEV statistics, which can be easily adapted or ex-
mutually uncorrelated, and all available statistical data have been tended to include more PEV details and different types of PEV
sampled. Thus, the obtained sampling results from simulation statistics to support specific case studies.
SUN et al.: NOVEL MARKOV-BASED TEMPORAL-SOC ANALYSIS FOR CHARACTERIZING PEV CHARGING DEMAND 163

Fig. 10. Adopted IEEE 53-bus test feeder with PEV penetration.
Fig. 11. Typical PEV state and SoC curves for (a): C-PEV; (b): O-PEV.

TABLE II
PEV STATISTICS OF BEIJING USED IN THE EXPERIMENT
PEV travel behaviors and adopted charging scheduling strategy.
Thus, two PEV charging scheduling strategies (i.e., “uncon-
PEV type M1 N1 trolled charging” and “smart charging”) studied in this paper
Market share ms msM 1 = 87.2% msN 1 = 12.8% are assumed to follow different rules to determine the start time
Proportion of PEV usage C-PEV O-PEV tsc of PEV charging actions.
rC = 82.7% rO = 17.3% The PEVs with “uncontrolled” charging strategy start normal
PEV travel times mean standard deviation
tm d of C-PEVs μ m d = 7 : 34 σ m d = 2.37 charging immediately upon the return of the last daily trips, as
tm a of C-PEVs μ m a = 8 : 20 σ m a = 1.86 given in (17), where j is the number of PEV samples:
te d of C-PEVs μ e d = 17 : 14 σ e d = 1.80
te a of C-PEVs μ e a = 18 : 03 σ e a = 2.34 tsc (j) = ta (j). (17)

Unlike the uncontrolled strategy that the PEV starts charging


upon its arrival, the “smart” charging strategy is able to shift the
IV. POWER FLOW ANALYSIS OF PEV CHARGING DEMAND
charging demand from grid demand peak time to valley hours.
This section adopts this model approach to quantify the im- The proposed modeling approach can obtain the probabilistic
pact of two PEV normal charging scheduling strategies (“un- correlation between multiple PEV states and SoC in individual
controlled” and “smart,” respectively). The study is carried out time slots, and the smart charging control can be integrated as
based on an IEEE 53-bus test feeder with the network bus and one of the management functionalities in the distribution net-
branch parameters obtained from [21], as shown in Fig. 10. work dispatch center. In details, the smart charging management
In addition to the parameters presented in Section II, the PEV can be implemented as follows:
field statistics and baseline demand in Beijing are obtained from 1) The expected number of PEVs with normal charging re-
[17] and [20] and scaled down proportionally in simulations, re- quests in the following time slots can be estimated at the
spectively, as given in Table II. Based on the PEV statistics, the dispatch center based on the Markov model and the PEV
PEV travel behaviors are simulated, and hence the PEV charg- state probability distribution (see Section-III).
ing demand can be estimated based on the proposed modeling 2) The SoC distribution (SoCt10 ,SoCt20 ,SoCt30 ) of these re-
approach (see Section III). The typical dynamics of PEV states quested PEVs in individual time slots can be estimated
and SoC over a day are shown in Fig. 11 for C-PEV and O-PEV, using the confusion matrix of HMM.
respectively, and hence the PEV travel cycles can be observed 3) The dispatch center schedules the start times of PEV nor-
in details. mal charging following a simplified rule: the start charg-
The start time tsc of PEV normal charging can directly af- ing times with the SoC belong to SoCt10 ,SoCt20 , and SoCt30
fect the temporal characteristics of PEV normal charging de- are postponed for 2, 3.5, and 5 h, respectively.
mand, assuming the normal charging process is uninterruptable. Base on the proposed modeling approach, the smart strat-
The normal charging start time of a PEV is determined by the egy can explicitly estimate the expected number of PEVs with
164 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 1, JANUARY 2018

TABLE III
NODE VOLTAGE PROFILE COMPARISON (P.U.)

Bus ID Uncontrolled strategy Smart strategy

Mean Standard Worst Mean Standard Worst


deviation case deviation case

8 0.996 0.039 0.928 0.995 0.031 0.951


9 0.986 0.062 0.874 0.986 0.048 0.917
12 0.976 0.071 0.846 0.976 0.055 0.896
18 1.047 0.003 1.041 1.046 0.002 1.043
22 1.015 0.030 0.963 1.014 0.024 0.980
30 1.037 0.013 1.015 1.037 0.011 1.022
33 1.020 0.027 0.972 1.020 0.021 0.990
39 1.027 0.016 0.999 1.027 0.013 1.009
40 1.039 0.013 1.016 1.038 0.010 1.024

normal charging requests and the SoC distribution of these


PEVs. Here, it can be calculated based on the available statistics
in this case study that the approximate average consumed time
for PEV charging of 20% SoC is about 1.7 h and the average
time that a PEV stays at home is 13.39 h. This indicates that the
smart strategy can meet the PEV charging requirement as well
as can realize the peak shaving and valley filling of distribution
network demand.
Now, we evaluate and compare the impact of these two
charging scheduling strategies on the network operation through
power flow analysis. In simulation, the expected total PEV nor-
mal charging demand of two strategies with 103 PEVs can be
obtained, respectively, as shown in Fig. 12(a). Fig. 12(b) shows
the total power loads based on the typical daily load curve with
different PEV penetration levels. Fig. 12(c) and Table III shows
the voltage profiles of the buses (with the charging facility) on
the test feeders with the normal charging demand of 103 PEVs.
The rate of peak-valley ξ is used to describe the peak-valley
gap, as expressed in
Pm ax − Pm in
ξ= × 100%. (18)
Pm in

In the case of uncontrolled charging strategy, it can be ob-


served that the peak-valley gap of the network bus voltage be-
comes more significant (ξ increases from 60.56% to 81.32%)
along with the increase of PEV penetration, i.e., from 103 PEVs
to 104 PEVs, and the variation of bus voltage becomes more
obvious (e.g., the lowest voltage magnitude and the maximum
difference over a day are 0.846 and 0.194 p.u. on bus No. 33,
respectively). From the network operation perspective, the in-
crease of voltage peak-valley gap can result in increased power
loss and potential violation of network constraints of bus volt-
age and line current. In particular, in the urban areas with large
residential demand, the uncontrolled charging behaviors of a
massive number of PEVs can bring direct adverse impact on the
power supply and electric transformers, and hence affects the
safe and stable operation of power systems. Therefore, the PEV
charging behaviors need to be appropriately managed.
The smart strategy can shift the PEV normal charging
Fig. 12. Comparison of two normal charging strategies: (a) total PEV
demand from grid peak period to valley hours (peak time of
normal charging demand; (b) total power loads; (c) bus voltage with PEV PEV normal charging changes from 21:30 to 2:00), as shown
integration. in Fig. 12(a). The reduction of the voltage peak-valley gap
SUN et al.: NOVEL MARKOV-BASED TEMPORAL-SOC ANALYSIS FOR CHARACTERIZING PEV CHARGING DEMAND 165

(ξ is reduced from 60.56% to 29.55%) and voltage fluctuation [3] U.S. Department of Transportation Federal Highway Administration,
(the lowest voltage magnitude and the maximum difference “Summary of travel trends: 2009 National household travel survey,” 2010.
[Online]. Available: http://nhts.ornl.gov/2009/pub/stt.pdf
over a day is 0.896 and 0.141 p.u. on bus 33, respectively) [4] T. Wu, Q. Yang, Z. Bao, and W. Yan, “Coordinated energy dispatching
can be observed in Fig. 12(b) and (c), respectively. It needs in microgrid with wind power generation and plug-in electric vehicles,”
to be highlighted here that this paper adopts a simple smart IEEE Trans. Smart Grid, vol. 4, no. 3, pp. 1453–1463, Sep. 2013.
[5] W. Yao, C. Y. Chung, F. Wen, M. Qin, and Y. Xue, “Scenario-based
charging strategy in the evaluation aiming to demonstrate the comprehensive expansion planning for distribution systems considering
effectiveness and benefit of the proposed charging demand integration of plug-in electric vehicles,” IEEE Trans. Power Syst., vol. 31,
modeling approach. In fact, more advanced and sophisticated no. 1, pp. 317–328, Jan. 2016.
[6] J. Taylor, J. W. Smith, and R. Dugan, “Distribution modeling requirements
charging strategies can be adopted to coordinate the charging for integration of PV, PEV, and storage in a smart grid environment,” in
actions of PEVs to achieve the operational objectives of Proc. IEEE Power Energy Soc. General Meeting, San Diego, CA, Jul.
distribution networks, e.g., our previous work [4] exploited the 2011, pp. 1–6.
[7] K. Qian, C. Zhou, M. Allan, and Y. Yuan, “Modeling of load demand due
coordinated PEV charging management solution considering to ev battery charging in distribution systems,” IEEE Trans. Power Syst.,
the presence of renewable distributed generators. vol. 26, no. 2, pp. 802–810, May 2011.
[8] E. Pashajavid and F. Shahnia, “Sensitivity of electric vehicles demand
profile to the batteries departure state-of-charge,” in Proc. Aus. Uni. Power
V. CONCLUSION AND FUTURE WORK Eng. Conf. (AUPEC), Perth, WA, Sep./Oct. 2014, pp. 1–6.
[9] T. J. Geiles and S. Islam, “Impact of PEV charging and rooftop PV pene-
This paper presented a generic Markov-based modeling ap- tration on distribution transformer life,” in Proc. IEEE Power Energy Soc.
proach for PEV travel behaviors and charging demand by the use General Meeting, Vancouver, BC, Jul. 2013, pp. 1–5.
of available PEV statistics and Monte Carlo simulation based on [10] J. R. Agüero, P. Chongfuangprinya, S. Shao, L. Xu, F. Jahanbakhsh, and
H. L. Willis, “Integration of plug-in electric vehicles and distributed energy
the temporal-SoC analysis of PEVs. Markov model and HMM resources on power distribution systems,” in Proc. IEEE Int. Elect. Veh.
were adopted to describe the average transition probabilities Conf., Greenville, SC, Mar. 2012, pp. 1–7.
among multiple PEV states and the probabilistic correlation be- [11] G. Li and X. P. Zhang, “Modeling of plug-in hybrid electric vehicle
charging demand in probabilistic power flow calculations,” IEEE Trans.
tween PEV states and SoC. Hence the temporal characteristics Smart Grid, vol. 3, no. 1, pp. 492–499, Mar. 2012.
of PEV states and charging demand in any given time slots [12] C. Kong, I. S. Bayram, and M. Devetsikiotis, “Revenue optimization
can be explicitly predicted and estimated, which will greatly frameworks for multi-class PEV charging stations,” in IEEE Access, vol. 3,
no., pp. 2140–2150, Nov. 2015.
help to optimally dispatch PEV charging demand or plan PEV [13] B. J. Johnson, M. R. Starke, and A. D. Dimitrovski, “Examining the
parking lots and charging stations in different geographical ar- potential impact of plug-in electric vehicles on residential sector power
eas. In addition, in order to demonstrate the effectiveness and demand,” in Proc. IEEE Power Energy Soc. General Meeting, Denver,
CO, Jul. 2015, pp. 1–5.
benefit of the proposed modeling approach, two PEV normal [14] O. Sundström, O. Corradi, and C. Binding, “Toward electric vehicle trip
charging scheduling strategies were studied on an IEEE 53-bus prediction for a charging service provider,” in Proc. IEEE Int. Elect. Veh.
test feeder with field data to assess the impact of PEV normal Conf., Greenville, SC, Mar. 2012, pp. 1–6.
[15] A. Chiş, J. Lundén, and V. Koivunen, “Scheduling of plug-in electric vehi-
charging demand on distribution networks. cle battery charging with price prediction,” in Proc. IEEE/PES Innovative
In the future, a number of directions are considered worth fur- Smart Grid Technol. Eur., Lyngby, Oct. 2013, pp. 1–5.
ther research effort. This paper adopted Monte Carlo simulation [16] A. M. A. Haidar and K. M. Muttaqi, “Behavioral characterization of elec-
tric vehicle charging loads in a distribution power grid through modeling
technique for generation of PEV instances. In fact, the proposed of battery chargers,” IEEE Trans. Ind. Appl., vol. 52, no. 1, pp. 483–492,
solution needs to be further validated and improved through re- Jan./Feb. 2016.
alistic distribution network case studies using a massive number [17] Beijing Transportation Research Center, “2011 Annual Report on
traffic development in Beijing,” Aug. 2011. [online]. Available: http://
of field statistics, e.g., PEV travel patterns, battery parameters; www.bjtrc.org.cn/InfoCenter/NewsAttach/aeb7c878-d31e-4f08-982f-
considering the increasing penetration of PEVs in urban areas, 3c17c717c87b.pdf
the proposed PEV demand modeling can be adopted in network [18] Y. Mu, J. Wu, N. Jenkinsa, H. Jia, and C. Wang, “A spatial–temporal
model for grid impact analysis of plug-in electric vehicles,” Appl. Energy,
planning tools to address the optimal placement and capacity vol. 114, pp. 456–465, Feb. 2014.
planning of parking lots and charging stations in distribution [19] Transport Research & Innovation Portal, “Mobile energy resources in grids
networks; the proposed model can be further extended to exploit of electricity,” 2010. [Online]. Available: http://www.transport-research.
info/sites/default/files/project/documents/20140203_154622_76425_
the scheduling of PEV charging and discharging behaviors un- Deliverable_2.1_Modelling_Electric_Storage_devices_for_Electric_
der various network operational conditions to achieve efficient Vehicles.pdf
demand side management, e.g., considering the presence of re- [20] L. Feng, “Electric vehicle charging station planning,” Ph.D. Dissertation,
School Elect. Eng. Automation, Tianjin Univ., Tianjin, China, 2013.
newable distributed generation and real-time electricity pricing. [21] R. Romero, J. F. Franco, and F. B. Leão, “A new mathematical model
The research outcome will be provided in future publications. for the restoration problem in balanced radial distribution systems,” IEEE
Trans. Power Syst., vol. 31, no. 2, pp. 1259–1268, Mar. 2016.
[22] J. Li, L. Ye, Y. Zeng, and H. Wei, “A scenario-based robust transmission
REFERENCES network expansion planning method for consideration of wind power
[1] J. C. Mukherjee and A. Gupta, “Distributed charge scheduling of plug-in uncertainties,” CSEE J. Power Energy Syst., vol. 2, no. 1, pp. 11–18,
electric vehicles using inter-aggregator collaboration,” IEEE Trans. Smart Mar. 2016.
Grid, vol. 8, no. 1, pp. 331–341, Jan. 2017. [23] C. Jungemann, S. Yamaguchi, and H. Goto, “Convergence estimation
[2] M. Ismail, I. S. Bayram, M. Abdallah, E. Serpedin, and K. Qaraqe, “Op- for stationary ensemble Monte Carlo simulations,” Int. Conf. Simulation
timal planning of fast PEV charging facilities,” First Workshop on Smart of Semiconductor Processes and Devices, 1997. SISPAD’97, Cambridge,
Grid and Renewable Energy (SGRE), Doha, Mar. 2015, pp. 1–6. MA, USA, Feb. 1997, pp. 209–212.
166 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 14, NO. 1, JANUARY 2018

Siyang Sun received the B.S. degree in automa- Wenjun Yan received the M.Sc. and Ph.D. de-
tion from Zhejiang University, Hangzhou, China, grees in control engineering both from Zhejiang
in 2014, where he is currently working toward University, Hangzhou, China, in 1991 and 1994,
the Ph.D. degree in electrical engineering. respectively.
His current research interests include electric He is currently a Full Professor at the Col-
vehicles grid integration, charging station plan- lege of Electrical Engineering, Zhejiang Uni-
ning, and demand response. versity. His research interests include robust
control, fault diagnose for hybrid systems, home
service robotics, and coordinated dispatch for
electricity and heat in microgrid.

Qiang Yang (M’09) received the Ph.D.


degree in electronic engineering and computer
science from Queen Mary, University of London,
London, U.K., in 2007.
He was with the Department of Electrical and
Electronic Engineering, Imperial College Lon-
don, London, U.K., from 2007 to 2010. He visited
the University of British Columbia and University
of Victoria Canada as a visiting scholar in 2015
and 2016. He is currently an Associate Profes-
sor at the College of Electrical Engineering, Zhe-
jiang University, Hangzhou, China, and has published more than 120
technical papers, applied 40 national patents, coauthored 2 books, and
several book chapters. His research interests include communication
networks, smart energy systems, and large-scale complex network mod-
eling, control, and optimization.
Dr. Yang is the member IET and the Senior Member of China
Computer Federation.

Vous aimerez peut-être aussi