Journal of Power Sources 267 (2014) 584e597

Contents lists available at ScienceDirect

Journal of Power Sources

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

Coordinating plug-in electric vehicle charging with electric grid:

Valley filling and target load following
Li Zhang, Faryar Jabbari, Tim Brown, Scott Samuelsen*
Advanced Power and Energy Program (APEP), University of California, Irvine, Irvine, CA 92697-3550, United States

h i g h l i g h t s

Proposed protocol effectively fills the overnight valley.

Computation and communication efforts are very modest.

Modified protocol can approach a desired target load.

a r t i c l e i n f o a b s t r a c t

Article history: Plug-in electric vehicles (PEVs) shift energy consumption from petroleum to electricity for the personal
Received 2 December 2013 transportation sector. This work proposes a decentralized charging protocol for PEVs with grid operators
Received in revised form updating the cost signal. Each PEV calculates its own optimal charging profile only once based on the cost
13 April 2014
signal, after it is plugged in, and sends the result back to the grid operators. Grid operators only need to
Accepted 16 April 2014
aggregate charging profiles and update the load and cost. The existing PEV characteristics, national
Available online 14 May 2014
household travel survey (NHTS), California Independent System Operator (CAISO) demand, and estimates
for future renewable generation in California are used to simulate PEV operation, PEV charging profiles,
Keywords:
Plug-in electric vehicle
grid demand, and grid net load (demand minus renewable). Results show the proposed protocol has
Grid operation good performance for overnight net load valley filling if the costs to be minimized are proportional to the
PEV charging net load. Annual results are shown in terms of overnight load variation and comparisons are made with
Decentralized control grid level valley filling results. Further, a target load can be approached in the same manner by using the
Valley filling gap between current load and the target load as the cost. The communication effort involved is quite
modest.

2014 Elsevier B.V. All rights reserved.

1. Introduction vehicle supply equipment (EVSE) [8,9]. The state and local gov-
ernments, as well as automakers, have shown interest in building a
Plug-in hybrid electric vehicles (PHEVs) and battery electric sufficient charging network. Previous work has presented analysis
vehicles (BEVs) are typically classified under the category of plug-in of the allocation of charging infrastructure [4,9]. There, it is shown
electric vehicles (PEVs) [1]. PEVs have drawn interest from gov- that with large PEV penetration, even with a reliable charging
ernment, automakers, and the public due to the potential to reduce network, the majority of the charging activities occur at home with
fossil fuel consumption, tailpipe emissions, overall greenhouse gas the current PEV characteristics and charging rates, due to the cheap
emissions, and operating cost [2]. A variety of research papers have night time residential electricity and the long dwelling time
evaluated PEV benefits quantitatively [3e6]. The California needed. Furthermore, charging time strategy has been showed to
Advanced Clean Cars programs mandates 1.4 million zero-emission have the most significant impact on charging cost reduction and
and PHEVs in California by 2025 [7]. However, a consensus has been overall grid operation. Here we focus on the details of coordinating
reached that one of the hurdles for large deployment (or accep- PEV charging, at home, with the electric grid.
tance) of PEVs is the shortage of charging infrastructure or electric The electricity demand and generation of the grid have to be
balanced at all times to assure operational stability. Charging PEVs
increases the electric demand and has the potential to change the
* Corresponding author. demand curve, if PEV penetration becomes significant. The time
E-mail addresses: [email protected], [email protected] (S. Samuelsen). needed to charge PEVs, for most travel demands, is less than the

http://dx.doi.org/10.1016/j.jpowsour.2014.04.078
0378-7753/
2014 Elsevier B.V. All rights reserved.
 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597 585

Notation L(ti) final load with PEVs charging

Tk time when cost is updated
ti time slot i in the 48-h window, e.g., 12 ame1 am, 1 am Vk vehicle number when cost updated
e2 am,., 11 pme12 am Tstep time interval for cost function updating
i time slot number, e.g., 1,2,., 48 Vstep vehicle number interval for cost function updating
Dt time slot duration, e.g. 60 min (1 h) k kth step to update cost function
Dtn ðti Þ plugged in time in time slot i for vehicle n, known sk(ti) aggregated charging profile for step k
n PEV number Ck(ti) cost function for charging at step k
tan home arrival time after the last trip for PEV n R(ti) maximum overall charging power at each time slot,
E(ti) electric demand known
D(ti) electric net load X(ti) overall charging load at each time slot, decision
xn(ti) charging energy at each time slot for vehicle n, decision variable
variable TL(ti) target load
rn(ti) maximum charging energy at each time slot for vehicle TCk(ti) cost function for charging at step k with target load
n, known

dwelling time overnight. Unlike day time charging, overnight and iterative, approach. In each iteration, a ‘price’ structure is
charging can be flexible and can be managed so that, aggregated communicated to the fleet of PEVs, so that each vehicle can develop
with overall demand, it results in lower generation cost and an optimal (with respect to the broadcasted cost) charging profile.
emissions. Generally, constant (or flat) demand curves are consid- These profiles are sent back to the central communication node or
ered beneficial for cost and environmental consideration [10]. the grid operator (e.g., the ISO e Independent System Operator),
Typically, wind and solar generation are treated as negative de- who will aggregate the total demand, based on the individual
mand since the power cannot be controlled in the same way as profiles, and broadcast a new price. Under relatively minor as-
other forms of generation. So the net load, total demand minus sumptions, the algorithms have convergence proofs. While the
renewable generation, is targeted to be flat or at least slowly results are quite impressive, there are some challenges. Both
varying. The problem can be simply stated as obtaining a charging techniques require the total number of PEVs be available for
pattern so that the final net load curve has the least variation over participation in the iterations needed in the optimization ([15] has
an extended time horizon, given an original net load curve from the results for the asynchronous case as well). Such an iterative
grid and the total charging demand for numerous PEVs. approach might require significant communication if the number
It has become clear that if there is a significant penetration of of vehicles is large. More crucially, these techniques do not ensure
PEVs, some form of “smart” or scheduled charging protocol will be each PEV is charging at the maximum charging rate, which is how
needed. The power requirement of a large number of PEVs at peak PEVs are currently charged. Ref. [16] attempts to address the last
or near peak times can lead to significant challenges in cost, de- concern by relying on a stochastic approach in which the start of
livery through grid, and even in generation and ramping capacities. the charging period is the decision variable in the optimization
This has led to several approaches to address this problem. problem, given the charging rate and SOC e which yields the
Generally, the main goal is to schedule and shift the charging de- charging duration. Under mild assumptions, the proposed iterative
mand of the PEVs to the late evening and very early morning when algorithm converges with probability one. Papers [17] and [18]
the overall demand is the lowest. These are often called ‘valley propose decentralized charging controls for PHEV to avoid trans-
filling’ approaches since they are aimed at leveling the overall de- former overloading, but cannot fill the overnight demand valley.
mand to reduce the need for shutting down and restarting of large Paper [19] utilizes system-wide or nodal price for PHEV in the
power plants. Of course, depending on specific, and relatively un- distribution network, however, it requires non-convex optimiza-
certain, costs associated with ramping and other considerations, it tion solving technique.
is possible that valley filling is not the optimal solution. For In this paper, we focus on the similar problem with somewhat a
example, results in Ref. [11] show that under certain combinations different tack. We propose two approaches that ensure charging
of level 2 charging, station penetration, and costs assigned to ramp occurs at the maximum power, as is required with the current
rates, etc., one can design a more desirable (e.g., less costly) charging technology (1.44 kW for level 1 and 3.3 kW for level 2
charging profile, though how such a global plan can be realized is EVSE), and the partial charging rate will lead to efficiency drop of
unclear. Here, we focus on the decentralized approach to address the converter [20]. As a key contribution, we attempt to minimize
this challenge, as centralized approaches are difficult to implement the amount of communication needed between the large fleet of
and unlikely to be accepted. PEVs and grid operator and do not require availability of all PEVs for
Among the decentralized approaches that have appeared initiating the charging time assignments. Similar to other ap-
recently, paper [12] first solves a centralized optimization problem proaches, it can be modified to address excessive ramp rates or
that takes into account costs associated with CO2 generation and/or possible intermittent renewable sources (with some reasonable
other economic and environmental costs. Based on the obtained prediction window).
average charging power, it then develops an algorithm that yields a The basic approach can be summarized as follows. We use a cost
decentralized implementation. Papers [13,14] use non-cooperative schedule that reflects the desirable ‘valley’ or ‘valleys’ for the PEVs
game concepts to develop a global valley filling protocol, under the to charge (by assigning low costs to such periods). This is shared
assumption that all BEVs have similar state of charge (SOC) and with individual vehicles, each solving a simple linear program to
other properties, and are plugged in at the same time. Paper [15] identify the periods for charging (at peak power), which will be the
removes the homogeneity assumption and allows varied SOC, lowest ‘cost’ e and overall demand e periods. The solution is then
max charge rate, etc. The approaches in Refs. [13e15] are aimed at sent back to grid operator for updating the charge structure. Note
solving the global valley filling problem through a decentralized, that this is not an iterative technique e there is only one set of data
586 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597

transmitted each way, once. The approach uses the natural flow of Table 1
PEVs in and out of the overnight (or valley) period, plus an enforced Simulation parameters for PHEVs and BEVs.

queue, and lets each vehicle sign up for a specific time period for Vehicle MPG kWh/mi All-electric Efficiency from AC charging
charging. By controlling the queue (and thus the rate of commu- type (DC) range grid to battery power (kW)
nication), one can ensure the lowest cost periods (grid level valleys) (miles)

are filled without solving the global valley filling problem directly. PHEVs 40 0.34 40 0.85 3.3
BEVs N/A 0.31 60 0.85 3.3

1.1. NHTS

The vehicle travel behavior data used in this paper are derived current production vehicles [22]. Gasoline price is assumed to be
from the 2009 National Household Travel Survey (NHTS) [21]. high enough throughout this work in order to ensure PHEVs use
Several processing steps were required in order to prepare the data. electricity first rather than gasoline. To simulate the future scenario,
In particular, data for California were selected, trips occurring this work uses a 10% penetration of all passenger cars (PC) and light
without a personally owned vehicle were deleted, person-chain duty trucks (around 2.1 million PEVs in California) [23]. So the
data were converted to vehicle-chain data, daily trips data with scaling factor from NHTS data to more than 2 million PEVs will be
unlinked destinations or significant over-speed were deleted, and around 100.
tours were organized into home based daily tours (first trip from
home, last trip to home) [9]. A total of 20,295 vehicles were selected 1.3. Renewables and net load
covering 83,005 single trips, with an average of 7.85 miles per trip
and 32.13 miles per vehicle, per day. The electricity sector in many countries and states have targets
Fig. 1 shows the histogram of the last home arrival time of the for meeting increasing fractions of their load demand with
day, with 15 min (0.25 h) intervals for all those 20,295 vehicles. The renewable resources to promote a shift towards a low-carbon, low-
peak arrivals occur in the late afternoon with almost 800 vehicles pollutant emission grid mix [10]. In California, the target is to
for the data set, which is around 4% of the total vehicles for this provide 33% of all retail sales of electricity from renewable re-
interval (around 17:00). The accumulation curve shows by 20:00, sources by the year 2020. Other states in the U.S. also aim to reach
80% vehicles have arrived and will stay until the next morning. similar goals to some extent.
Fig. 1 also shows when vehicles leave home for the first trip of Fig. 2 shows the electric demand E(ti) and net load D(ti) based on
the day and the peak occurs in the early morning at around 7:00, by wind and solar installed capacities at around 30% renewable
which time, 80% of vehicles have not left. Combined with the arrival penetration in terms of hourly resolved and monthly averaged
time, shown in the ‘Plugged in’ curve, 80% of vehicles can be co- signals. Those capacities were specified by California Public Utilities
ordinated with the grid at home for almost half of the day, from Commission (CPUC) [24]. It is well known that the demand has a
20:00 to 7:00 in the next day. So this time period is considered to be diurnal pattern with a valley hours after midnight. The same
the most effective coordinating window between PEVs and grid. pattern also can be observed for the net load for most of the days.
However, as shown from day 98e99 and day 108e109, additional
1.2. Vehicle information large peaks and valleys may also exist, in particular the relatively
large valley in the afternoon. Ideally, the aggregated PEV charging
Similar to other research [11e16], this study focuses on the gird profile can be used to smooth this curve as much as possible so the
impact of macro scale of vehicle behavior where the detailed final load is met with lower cost and/or emissions.
physical vehicle model was not considered; instead a parameter- To fulfill this goal, in the rest of this paper, the optimization of
ized vehicle operation and charging model was used. Table 1 shows valley filling from the grid perspective will be first introduced to
vehicle parameters used in this study which were all derived from provide a reference solution. And it will introduce the mechanism

Fig. 1. Vehicles home departure and arrival time distribution and charging availability.
 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597 587

x 10 x 10
3.5 4.5
Net load with 30% Renewables Net load with 30% Renewables
Electric Demand 2005 Electric Demand 2005
4
3

Load Demand (MW)

Load Demand (MW)
3.5
2.5

2
2.5

1.5
2

1 1.5
98 100 102 104 106 108 110 0 1 2 3 4 5 6 7 8 9 10 11 12
Time of the Year (Day) Time of the Year (Month)

Fig. 2. Hourly resolved electric demand and net load for ten days (left) and on a monthly average basis (right).

of optimal charging from the perspective of each individual PEV as where X(ti) is the overall charging power at each time slot ti, and B is
well as the protocol to update the cost function broadcast by the the total charging energy of all PEVs for a whole day, which is
grid operator to achieve the optimality at the grid scale. Addition- considered to be known, which the summation of the charging
ally, the year-long results will be shown and compared to the energy bn for individual PEV. This problem format or the corre-
original net load and the results from grid level valley filling. Finally, sponding result has been seen in papers [11,13,15,25]. The well-
a modified protocol will be introduced for the final load to approach known solution is the following
a pre-defined target load which may not only be the solution of grid
level valley filling. Xðti Þ ¼ ðl  Dðti ÞÞ
þ
¼ maxfðl  Dðti ÞÞ; 0g

X þ
2. Grid level valley filling ðl  Dðti ÞÞ ¼ B (3)
i
As mentioned earlier, the grid favors certain types of load
curves, independent of the behavior of individual PEVs. Valley where l is the height to which that the valley is filled.
filling is thus a strong preference, in which ideal valley filling is to However, unless all vehicles are plugged in for the whole time
solve a convex optimization problem with the constraints on total horizon, this solution ignores an important constraint associated
energy consumption of PEVs. The optimization details can be with the overall charging power: the ideal valley filling result is not
formatted as follows. feasible if there are not enough PEVs plugged-in for a specific time
slot according to the PEV charging availability shown in Fig. 1.
X
min ðDðti Þ þ Xðti ÞÞ2 (1) Similarly, In Refs. [11], a Charge Flexibility Constraint (CFC), is added
i to consider the overall charging power constraint, as well. However,
the Charge Flexibility Constraint indicates a sharp decrease on the
Subject to maximum charging power (plugged in PEVs) from 1:00 to 6:00,
X X which is quite different than Fig. 1. In this time period, most PEVs are
Dt  Xðti Þ ¼ B ¼ bn (2) at home, thus the maximum charging power should be relatively
i n
flat. Here, Fig. 1 is used to derive an additional constraint shown in
Equation (4) to provide an upper bound on the overall charging

4
x 10

4 Net Load without PEVs

Net Load with Global Valley Filling, Unconstrained Power
3.5 Net Load with Global Valley Filling, Constrained Power
Net Load with Immediate Charging
Load Demand (MW)

Charging Profile from Global Valley Filling, Unconstrained Power

3
Charging Profile from Global Valley Filling, Constrained Power
Upper Bound on Charging Power
2.5

1.5

0.5

0
0 6 12 18 24 30 36 42 48 54 60 66 72
Day 97 to Day 99 (Hours)
Fig. 3. Comparison of ideal valley filling results without power constraint and constrained valley filling results with power constraint.
588 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597

power, depicted by the black line marked with the cross sign in 3.1. Individual PEV charging
Fig. 3. Essentially, R(ti) is the product of the amount of plugged in
PEVs and the individual charging power (3.3 kW in this study). This The initial charging strategy used for individual PEVs is similar
constraint is independent from the electric load curve and in reality to that in Ref. [9], except that here we are focused on overnight (at
can be derived from historical plug-in and plug-out data. As contrast home) period. In summary, given charging vehicle demand (i.e.
to the ideal valley filling without a power limit, the combination of energy), charging constraints (i.e. plugged in time window and
(1) (2) (4) can be defined as constrained valley filling. charging power limits), charging cost (i.e. price of electricity, as
provided by grid operator), each PEV finds the optimal way to
Xðti Þ  Rðti Þ (4) charge such that its individual cost can be minimized, as shown
Fig. 3 shows an example of the optimal solutions for both cases below.
on the net loads from day 97 to day 99 (hour 0 to hour 72). It is Problem formulation
assumed that only the PEVs that arrive in Day 97 and 98 count. Day
98 is not a typical day since the largest valley does not occur 1. Decision variables:
overnight but in the early afternoon (at hour 39 on the graph). The
ideal valley filling result, the dotted line, requires more vehicles The SOC increase (or electricity recharged) of the PEV n for each
than are actually plugged in between hours 36e40 (i.e. arrived and time slot ti given by xn(ti).
plugged in). For another period from hour 1 to hour 4 on the graph,
when there is not a single PEV since no vehicle has arrived, the ideal 2. Cost function:
valley filling draws power from the grid as well. When the The summation of the total charging cost is given by:
inequality constraint in Equation (4) which is represented by the X
Cðti Þ  xn ðti Þ (5)
black curve marked with the cross sign at the bottom of the figure is
i
used, no power flows to PEVs before hour 6. And, the deepest valley
from hour 36 to 42 ends up being filled only partially, as shown by where C(ti) is the cost per kWh (DC) during the time ti, derived from
the solid line. Although the situation of having the biggest valley in the electricity cost provided by grid operator.
the afternoon may not happen often, it underlines the point that
the ability of PEVs to alter electric load significantly may be 3. Constraints:
somewhat limited to the overnight period. Although constraint (4) 1) The total energy charges during the home dwelling period
considers PEV availability, it is most restrictive in the early after- should match the consumed energy during the day (known
noon periods. During late night periods, with some of the PEVs fully and fixed):
charged, the real available PEV number becomes smaller than that
X
of plugged in PEVs. Thus, during these periods, (4) overestimates xn ðti Þ ¼ bn (6)
the limit and is not a restrictive constraint. Fig. 3 also shows the i
result from the immediate charging as a benchmark, depicted by
the dashed line. It increases the peaking load by 2 GW for day 97
and day 98, indicating the necessity of the coordinated charging 2) The lower bound on xn(ti) is zero and for the upper bound, it
control, such as the valley filling approach. is the product of the charging power at home pn(ti) (3.3 kW),
Naturally, a perfect valley filling solution from (1) (2) (4) may plugged in Dtn ðti Þ time for charging in each time slot Dtn ðti Þ;
not be achievable depending on the specific shape of the net load, and charging efficiency h (0.85).
the availability of the PEVs, and the individual level charge needs. A
complete global valley filling solution would require a problem 0  xn ðti Þ  rn ðti Þ ¼ pn ðti Þ  Dtn ðti Þ  h (7)
with decision variables on the order of the PEV population, while
the ideal valley filling requires a number that is on the order of time
slots (as it solves the aggregate charge at each time slot). As a If a PEV arrives home at 17:30 and leaves at 7:45 next day, and the
compromise, in this work, the optimal solution of the constrained time slot duration Dt is 1 h (60 min), then the plugged in time for
valley filling, i.e., (1) (2) (4), is considered the reference profile. For charging from 17:00 to 18:00, Dtðt18 Þ; is 0.5 and the plugged in time
simplicity, the constrained valley filling will be called valley filing in for charging from 7:00 to 8:00 in the next day, Dtðt32 Þ; is 0.75.
the rest of this paper.
4. Assumptions on the variables:
1) The span of plugged in time for each vehicle is fixed by the
3. Protocol of individual PEV charging and cost updating NTHS data. That is, each PEV controller knows the time for
the first trip the next day.
The charging profile of individual PEVs is trivial compared to 2) The electricity cost is not exactly the same in any two time
the grid load and only the aggregated profile has to be considered. intervals, i.e., C(ti) values are distinct.
In a decentralized scheme, every vehicle makes the best decision 3) The charging power and AC to DC efficiency are assumed to
for its own, in terms of the overall electricity cost available to it. be a constant and known.
This section will introduce the mechanism of optimal charging for
single PEVs as well as the protocol to update the cost function
broadcast by the grid operator to achieve the optimality at the grid
scale, to the extent possible (given the constraint on PEV avail- 3.1.1. Charging profile
ability). After that, a one-day result will be shown at different The key is the price made available to each PEV. At specific in-
updating frequencies. Then, the communication effort will be stants (e.g., every 30 min or after every 10,000 cars have ‘registered’
evaluated to verify the feasibility of the proposed protocol. Finally, for the night), the grid operator sends the recently arrived vehicles
the year-long results will be demonstrated and compared to the a price profile C(ti). We start with the simplest option: C(ti) is simply
original net load and the results from valley filling formatted in the electricity net load (i.e. demand minus renewable) D(ti). In re-
the previous section. ality, the cost must have different values at different times.
 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597 589

1 ‘bump’ in price around 3 a.m. Note that this process does not pre-
Normalized Charging Cost

Charging Cost from Net Load vent intermittent charging profiles, which can happen if the
0.9
apparent cost to the PEV has multiple local (and similar in depth)
0.8 peaks and valleys.
The intermittent charging is considered to be a milder charging
0.7 condition compared to continuous charging. In this pattern, battery
temperature will be lower which has less impact on degradation,
0.6
since temperature is the leading factor for degradation [26e28].
0.5 Similarly, intermittency will allow the converter and other elec-
10 15 20 25 30 35 tronics to cool down. If continuous charging has to be guaranteed,
Time of Two Days (Hour)
an alternative algorithm can be the following: 1), each PEV calcu-
5 lates its required charging time at arrival; 2), depending on the
Charging Profile, PEV1 departure time in the next day, the PEV finds out all the feasible
4 Charging Profile, PEV2 continuous charging windows (i.e. charging starting points); 3),
according to the electricity cost broadcast, the PEV does a simple
Power (kW)

3
line search to determine the optimal charging window (i.e.
2 charging starting point). Assuming a continuous window that ac-
commodates the needed charge exists, it is easy to see that the
1 algorithm leads to a continuous block.
Finally, note that the communication between grid operator and
0
10 15 20 25 30 35 PEV is quite modest: one set of prices from grid operator, one set of
Time of Two Days (Hour) charging values for PEV, once.
Fig. 4. Charging profiles for two PEVs with different energy requirements and plug-in
window. 3.2. Cost updating

Given that individual PEVs choose the lowest cost periods, the
Since the cost for a specific time slot is different than any other, question becomes how to control/update the cost function such
optimizing the cost function in (5) above will choose the lowest that the aggregated charging profile can be close to the optimum
point to charge and then jump to the second lowest one with the for grid operation. Intuitively, a simple approach would be based on
maximum charging power until the SOC reaches 100%. In this a cost C(ti), where the original net load curve D(ti) can be used as the
sense, the unrealistic partial rate charging [11e15] can be avoided. cost directly, as shown in the previous section. The basic scenario
Indeed, there can be at most one partial charge time period, which for cost updating is that the grid operator provides the net load
means the charging stops partly through that period (e.g., 15 min forecast D(ti) each day and all PEVs will respond to the same signal.
into the a 60 min period). The detailed proof is shown in the Fig. 5 shows the final result if the same net load is shared with
Appendix. the total population of PHEV40 (40 miles all electric range) for a
Fig. 4, lower panel, shows charging profiles for two PEVs by typical day. The charging load exhibits a large peak at 2:00 (26:00
assuming the charging cost on the top pane, which is obtained from on the graph) where the original net load has its minima. The in-
the net load from day 98 to day 99. The two PEVs arrive home at cremental load is over 7 GW and lasts a short time. In this sense, the
5:15 and 16:40, respectively. So PEV 1 decides to charge from 12:00 new peak will expose a significant extra burden on the grid.
to 16:00 and late in the evening to guarantee full charge. For PEV 2, As shown in the introduction, the PEVs arrival time is well
the lowest cost available occurs overnight so it decides to charge distributed so the new peak above was built up through a relatively
then with charging turned off for about 10 min, due to the small long time (more than half day) by the individual decisions made at

4
x 10
3.2 4
x 10
Net Load without PEVs 3.2
Final Load with Valley Filling Net Load without PEVs
3 Final Load with Valley Filling
Final Load with PEVs Protocol 3
Cost Updated at 16:00
Load Demand (MW)

2.8 Final Load with PEVs Protocol

Load Demand (MW)

2.8

2.6
2.6

2.4
2.4

2.2
2.2

2
2

1.8
5 10 15 20 25 30 35 40 45 1.8
5 10 15 20 25 30 35 40 45
Time of Two Days (Hour)
Time of Two Days (Hour)
Fig. 5. Update cost function every 24 h with PHEV40 charging at 3.3 kW.
Fig. 6. Update cost function every 12 h.
590 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597

x 10 Net Load without PEVs x 10 Net Load without PEVs

3.2 3.2
Final Load with Valley Filling Final Load with Valley Filling
Cost Updated at 10:00 Cost Updated at 8:00
3 Cost Updated at 16:00 3
Cost Updated at 12:00

Load Demand (MW)

Load Demand (MW)
Cost Updated at 22:00 Cost Updated at 16:00
2.8 Final Load with PEVs Protocol 2.8 Cost Updated at 20:00
Cost Updated at 24:00
2.6 2.6 Final Load with PEVs Protocol

2.4 2.4

2.2 2.2

2 2

1.8 1.8
5 10 15 20 25 30 35 40 45 5 10 15 20 25 30 35 40 45
Time of Two Days (Hour) Time of Two Days (Hour)

Fig. 7. Update cost function every 6 h (left) and 4 h (right).

different times. A potential solution to lower the peak is for the grid a. If fixed time interval is chosen, all of the individual profiles,
operator to change the cost function periodically so that PEVs generated from the PEVs that have arrival time, tan, in the
gradually fill the valleys in the updated cost function and avoid interval ½Tk1 ; Tk Þ are aggregated.
charging at the new peak. This process can be treated as a natural
valley filling. X .
sk1 ðti Þ ¼ xn ðti Þ hcn s:t: Tk1  tan < Tk (9)
Given that a 3.3 kW maximum load for one PEV is trivial n
compared to the entire grid load, the updating of the cost can be
accomplished such that a number of vehicles can register, obtain
prices, select charging times and communicate the results with grid Tk ¼ Tk1 þ Tstep ; T0 ¼ 4 am (10)
operator. A typical time interval Tstep, or a set number of PEVs Vstep,
or a set amount of the net load change in a time window (e.g., 23:00
to 6:00 in the next day) can be used to trigger the updating. Below, b. If the fixed PEVs amount is chosen, each aggregation takes
the protocol’s details and results will be shown based on the first place with the interval of vehicle number(Vk1, Vk).
two methods.
X .
sk1 ðti Þ ¼ xn ðti Þ hcn s:t: Vk1 < n < Vk (11)
1. Updating the cost function n
At each updating instant, the cost function is updated by
adding the aggregated PEV charging profiles to the previous cost
function. Vk ¼ Vk1 þ Vstep ; V0 ¼ 0 (12)

Ck ðti Þ ¼ Ck1 ðti Þ þ sk1 ðti Þ; C0 ðti Þ ¼ Dðti Þ (8) In either case, it is important to limit the number of PEVs that
will be aggregated between two cost broadcasts such that the load
increment is small enough to accomplish smoothing of the overall
load profile. This can be precisely controlled by Tstep or Vstep. In
2. Profile aggregating Section 3.3, the tradeoff between more frequent (more communi-
As discussed before, we study the mechanism for updating cation) and less frequent (less smooth load profile) updates will be
the cost: fixed time internals and fixed number of vehicles. discussed.

4 4
x 10 x 10
3.2 Cost Updated Every 30 Minutes
Net Load without PEVs 2.38 Cost Updated Every 30 Minutes
Final Load withValley Filling Net Load without PEVs
3 2.37
Final Load with PEVs Protocol Final Load with Valley Filling
Load Demand (MW)
Load Demand (MW)

2.36 Final Load with PEVs Protocol

2.8
2.35

2.34
2.6
2.33
2.4 2.32

2.31
2.2
2.3

2 2.29
2.28
1.8
5 10 15 20 25 30 35 40 45 21 22 23 24 25 26 27 28 29 30

Time of Two Days (Hour) Time of Two Days (Hour)

Fig. 8. Update cost function every 30 min.

 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597 591

x 10 x 10
3.2 Cost Updated Every 10,000 PEVs 2.37
Net Load without PEVs Cost Updated Every 10,000 PEVs
2.36 Net Load without PEVs
3 Final Load with Valley Filling
Final Load with PEVs Protocol 2.35 Final Load with Valley Filling

Load Demand (MW)

Load Demand (MW)
Final Load with PEVs Protocol
2.8 2.34

2.33
2.6
2.32

2.4 2.31

2.3
2.2 2.29

2.28
2
2.27

1.8
5 10 15 20 25 30 35 40 45 21 22 23 24 25 26 27 28 29
Time of Two Days (Hour) Time of Two Days (Hour)

Fig. 9. Update cost function every 10,000 PHEVs.

x 10 3.3 kW Charging x 10 6.6 kW Charging

3.2 3.2
Cost Updated Every 30 minutes Cost Updated Every 30 minutes
Net Load without PEVs Net Load without PEVs
3 3
Final Load with Valley Filling Final Load with Valley Filling
Load Demand (MW)

Load Demand (MW)

Final Load with PEVs Protocol Final Load with PEVs Protocol
2.8 2.8

2.6 2.6

2.4 2.4

2.2 2.2

2 2

1.8 1.8
5 10 15 20 25 30 35 40 45 5 10 15 20 25 30 35 40 45
Time of Two Days (Hour) Time of Two Days (Hour)

Fig. 10. Continuous charging with cost updated every 30 min for 3.3 kW (left) and 6.6 kW (right) charging power.

3.3. Protocol results and analysis fewer PEVs arrive home in the morning and the late evening while
more PEVs arrive home in the late afternoon and early evening.
3.3.1. One-day results That is why curves indicating costs are denser at the bottom and
Fig. 6 shows the results where Tstep is 720 min (12 h). The top, and less dense (i.e. larger change in each step) in the middle.
increment of the net load can be divided into two parts by the line Near the end of the process, i.e., very late evening, fewer PEV ar-
marked with the star sign, the one before 16:00 and the one after. rivals help smooth the demand curve.
Thus, the line shows the cost received by all the PEVs which arrive As shown in Fig. 8, when Tstep is reduced to 30 min the final
after 16:00, while the solid line was the cost received by those curve is quite flat. The detailed examination on the right hand side
arriving before 16:00. As expected, the first one has the same trend indicates the variation of the curve is less than 200 MW from 23:00
as observed in Fig. 5 and the second fills the two valleys to two to 4:00 in the next day. Thus, generators, especially load followers,
smaller peaks, depicted by the thick line. can have a more steady operating condition overnight.
Fig. 7 shows the resutls when Tstep are 360 and 240 min (6 h and Another observation is the creation, and later filling, of new
4 h) respectively. Each curve shows the load profile (cost function) peaks and valleys in the cost function curve in successive steps. This
updated at the corresponding time shown in the legend. First, the is rather intuitive for this strategy. For example, all PEVs arrived
final net load with PEV protocol becomes smoother with a between 16:00 and 16:30 are given the same cost function. They
decreased Tstep. Second, the difference between two costs will optimize their own cost, i.e. they will try to be on where the
Ck(ti)  Ck1(ti) is small at the beginning which gets larger and cost is lowest. After updating, the new cost would be higher at
eventually gets small again. This pattern is due to the distribution of those time slots and PEVs arriving between 16:30 and 17:00 will try
the vehicles arrival times in Fig. 1. With the same time interval, to avoid charging at those time slots.

Table 2
One-day communication effort of the proposed protocol.

Result type Update method Max arrival Min time interval to Max PEVs aggregated Update times
(PEVs per 15 min) update (minute) in one update per day

Analytical Fixed time (Tstep) 4%  N Tstep 4%  N/15  Tstep 1440/Tstep

(N, population) Fixed PEVs (Vstep) 4%  N Vstep/(4%  N/15) Vstep N/Vstep
Example Fixed time (Tstep) 84,000 30 167,000 48
(Tstep ¼ 30 min, Fixed PEVs (Vstep) 84,000 18 100,000 21
Vstep ¼ 100,000,
N ¼ 2.1 million)
592 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597

4
x 10 Net Load without PEVs
3.5
Final Load with Valley Filling
Final Load with PEVs Protocol
3 PEVs Charging Load from Protocol
PEVs Charging Load from Valley Filling
Upper Bound on PEVs Charging Power
Load Demand (MW)
2.5

1.5

0.5

0
105 108 110 113 115 118 120
Day of the Year
Fig. 11. Annual results from valley filling and proposed protocol with 30 min updating interval.

The fixed number of PEVs can also be used for triggering the the same where costs are updated by Tstep or Vstep. Fig. 10 shows the
updating of the cost (see Fig. 9) when it is preferred to control the results of continuous charging with costs updated every 30 min, as
variation of the cost in each step. Because the charging power is a comparison to the results of the main protocol with intermittent
fixed to be 3.3 kW, the maximum change between Ck(ti) and Ck1(ti) charging in Fig. 9. It can be observed that the final net load with
is limited by the product of charging power and the PEVs number PEVs is smooth, though not as flat as the one with intermittent
Vstep. In this approach, the grid operator develops a queue, as PEVs charging. There exists a bump in the middle of the overnight valley.
arrive, and after each Vstep vehicles, the cost is updated and shared This is due to the fact that the average charging time is over 3 h, for
with the next Vstep vehicles. In the example shown here, Vstep is set 3.3 kW power. Forcing continuous charging inevitably leads to a
to be 100,000 such that the maximum change on the cost can be sizable number if PEVs having overlapping charging time, resulting
limited to 330 MW step by step. Fig. 9 shows the results in com- in the shape seen in Fig. 10. Higher charging power would mitigate
parison with the first method. The final variation lies within a this issue. On the right hand side of Fig. 10, the results are shown
250 MW window from 23:00 to 4:00 in the next day. The cost with 6.6 kW charging power. As expected, the final load with PEVs
function increment is more consistent and limited by 300 MW as is as flat as the intermittent 3.3 kW charging in Fig. 8. However,
expected in the analysis above. 6.6 kW charging is not likely to be popular at least in the near future
As discussed earlier, continuous charging can also be guaranteed because it increases the component cost, makes the distribution
by changing the algorithm to a simple line search of the charging network less stable, and has minimal benefits in reducing the
start point for individual PEVs. The rest of the protocol can remain operating cost [9,29e31].

4
150 x 10
2.8
Net Load without PEVs
Current Load at 10:00
2.6 Current Load at 12:00
Current Load at 14:00
Load Demand (MW)

2.4 Current Load at 16:00

100 Current Load at 18:00
Counts (Days)

Final Load with Original PEVs Protocol

2.2

50
1.8

1.6

1.4
0 5 10 15 20 25 30 35 40 45
5 6 7 8 9 10 11 12 13 14
Window Width (Hours) Time of Two Days (Hour)
Fig. 12. Histogram of the maximum consecutive time with load varying less than Fig. 13. Final load of the original PEVs protocol for day 108 to day 109 and loads at
300 MW. different times.
 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597 593

4
4
Net Load without PEVs x 10 Net Load without PEVs
x 10 Target Load by Global Valley Filling 2.8
Final Load with Valley Filling
Target Load by Global Ramp Rate Reducing Current Load at 10:00
3
Modified Cost for Target Load by Global Valley Filling 2.6 Current Load at 12:00
Modified Cost for Target Load by Global Ramp Rate Reducing

Load Demand (MW)

Current Load at 14:00
2.5 2.4 Current Load at 16:00
Current Load at 18:00
Final Load with Modified PEVs Protocol
Load Demand (MW)

2 2.2

2
1.5

1.8
1
1.6

0.5
1.4
5 10 15 20 25 30 35 40 45

0 Time of Two Days (Hour)

Fig. 16. Results from modified protocol with time slots 11:00 to 18:00 prioritized by
-0.5 using the solution of valley filling as the target.
5 10 15 20 25 30 35 40 45
Time of Two Days (Hour)
Compared to methods in many existing publications [12e15],
Fig. 14. Two examples of the target loads, solution of valley filling and solution of ramp
rate reducing.
the calculation and communication efforts offer advantages as: 1)
one linear optimization calculated and one profile transmitted at
the PEV side without any global information; 2) no more than 10%
of profiles are aggregated in 30 min at the operator side; 3) the cost
3.3.2. Communication effort is updated no more than 50 times per day at the operator side; 4)
For the proposed protocol, either intermittent or continuous further, PEVs do not have to wait until a specific time to participate,
charging, each PEV receives the load curve (cost function), calcu- which increases the potential to change the net load in late
lates its profile, and transmits the results one time, while the grid morning and afternoon.
operator needs to receive all of the individual profiles in the set
before updating. So the calculation effort is decentralized to indi-
3.3.3. Annual results
vidual PEVs while the receiving and aggregating efforts are on the
To understand the performance of the protocol for different net
grid operator, proportional to the PEVs population N. In Table 2, the
load curves, the same PEV charging requirements repeated each
amount of individual profiles received in a step and the frequency
day with 30 min for Tstep has been implemented in the simulations
that the cost is updated are evaluated at the largest communication
for the whole year. As a reference, the valley filling technique has
burden. To obtain the results shown in Fig. 8, the cost is updated 48
been conducted separately. Fig. 11 is the snapshot of the results
times per day and no more than 167,000 PEV profiles are trans-
from day 105 to day 120. The charging profiles from two methods
mitted in 30 min. As for the results in Fig. 9, the minimum updating
are depicted at the bottom by the dotted and dashed lines. The line
interval is 18 min.
marked with the cross sign shows the upper bound of the charging
power, and the dashed and thick lines exhibit the final load with
PEV charging.
x 10
4
The two charging profiles (i.e., valley filling and the protocol
Net Load without PEVs
2.8 suggested here) are very close to each other especially when there
Target Load by Valley Filling
Current Load at 10:00 is only one deep valley overnight. The correlation coefficient of the
2.6 Current Load at 12:00 two profiles is 0.98 and the two resulting final loads exhibit less
than 0.02% difference on the objective function in Equation (1). The
Load Demand (MW)

Current Load at 14:00

2.4 Current Load at 16:00 only obvious difference occurs when there is another relatively
Current Load at 18:00 deep valley in the afternoon as shown in the results from day 108e
2.2
Final Load with Modified PEVs Protocol 109 or day 113e114. In that case, the valley filling provides a result
with two valleys filled up close to the same level while the pro-
2 posed protocol does not fill the first valley in the afternoon and fills
the overnight valley to a higher level. The main reason for that is
1.8
the overnight valley is relatively deep, leading to a smaller cost
function value so that PEVs participating in the protocol at the early
steps will choose to charge overnight. When it is filled to a higher
1.6
level than the afternoon, time has passed the first valley and the
rest of the PEVs can still only choose overnight to charge. It needs to
1.4
5 10 15 20 25 30 35 40 45 be noted that none of the existing research has addressed this since
Time of Two Days (Hour) they require that most PEVs arrived home to initiate the protocol,
often quite late in the day, e.g. midnight. Naturally, the afternoon
Fig. 15. Results from modified protocol without prioritizing by using the valley filling valley cannot be filled in practice. However, for the proposed pro-
solution as the target. tocol in this work, one possible solution is to come up with a final
594 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597

x 10 x 10 Net Load without PEVs

2.8 2.8
Target Load with Ramp Rate Reducting
Net Load without PEVs
Final Load with Modified PEVs Protocol (prioritizing)
Target Load with Ramp Rate Reducting 2.6
2.6
Final Load with Modified PEVs Protocol (prioritizing)

Load Demand (MW)

Load Demand (MW)
2.4 2.4

2.2 2.2

2 2

1.8 1.8

1.6 1.6

1.4 1.4
5 10 15 20 25 30 35 40 45 5 10 15 20 25 30 35 40 45

Time of Two Days (Hour) Time of Two Days (Hour)

Fig. 17. Two sets of Results from modified protocol with time slots 11:00 to 18:00 prioritized by using the solution of ramp rate reducing as the target.

net load curve (e.g., the valley filling results) as a target to approach that overnight valley is more attractive than the afternoon one at
and prioritize the afternoon valley by changing the cost function. the beginning. By the time the overnight valley is filled to the same
The details for this approach are discussed in Section 4. level as the afternoon one, it is too late for enough PEVs to fill the
Another important aspect, to evaluate the performance of the first one. In order to fill the first valley, or more precisely, to
proposed protocol, is the final net load variations for each night. approach a target load which the grid operator favors, here a
First, a maximum variation of the final load curve is defined (e.g., modified version is built as follows: 1) grid operators define a target
300 MW). Second, the length of the consecutive time window load by considering constraints at the grid level; 2) the cost
that meets this bound is found for every night of the year. This broadcast is changed to be the current load minus the target load;
window indicates for how long the balancing generators can 3) when appropriate, scaling is used to adjust the new cost function
operate at a relatively constant condition. Fig. 12 is the distribu- to ensure the desired overall charging profile. The local optimiza-
tion of the width of this time window by setting the maximum tions of individual PEVs along with the communication scheme
variation to 300 MW. For most of the nights, the window width is remain the same as before. These steps will be described in detail
above 7e8 h and there is only one night with a 5-h window and below, followed by daily results.
21 days with a 6-h window. The widest window reaches up to
13e14 h. The widely spread distribution is due to the variety of
the original net loads for different nights, shown in Fig. 11. A 4.1. Modified protocol
narrower and deeper net load valley tends to have a shorter
consecutive time of flat net demand, as day 118 shows, while a For the first step, a target load curve with PEVs charging, TL(ti) is
wider valley is likely to have a longer consecutive time of flat the summation of the known net load and the overall PEV charging
demand as day 119 shows. load X(ti), which is unknown but bounded according to Equation (2)
In summary, in the proposed protocol, each PEV calculates its and inequality (4).
own optimal profile based on the broadcast cost function which is
updated frequently. The charging load of PEVs can fill the overnight Þ ¼ DðtP
TLðti( i Þ þ Xðti Þ P
valley, often to a flat final net load curve. A valley in the afternoon Dt  Xðti Þ ¼ B ¼ bn (13)
imposes difficulty, however, as it lies outside of the most effective s:t: i n

charging zone, from 20:00 to 7:00 in the next day, when a large Xðti Þ  Rðti Þ
number of PEVs are plugged in. One example for the target load is to use the solution of valley
filling, formed by (1) (2) (4). The solution is depicted by the
4. Modified protocol for target load following dashed line in Fig. 14. Another example for the target load is
the solution of ramp rate reduction, which is aimed at smoothing
When there exists a relatively large valley (or multiple valleys) the final load for generation cost reduction [11]. In this study, the
in the afternoon due to large solar/wind generation (or intermittent objective function for ramp rate reduction is shown in Equation
solar/wind generation), it leads to the following potential scenario: (14), which minimizes the sum of squares of ramp rates between
power plants have to ramp up and down quickly to meet this two consecutive time slots. The total energy and maximum power
particular load, increasing cost and generating more emissions or, constraints associated with X(ti), in Equation (2) and (4), remain
alternatively, some of the solar/wind power is curtailed. It is the same. The ramp rate problem consisting of (14), (2) and (4), is
therefore preferable to fill or smooth the afternoon valley(s). As also a convex optimization problem and solvers, such as CVX and
shown in the left pane of Fig. 13, the original protocol is not aimed Quadratic Programming in MATLAB, can provide a solution fast
at accomplishing this task and does not fill the first valley at all. As and reliably. The solution to this problem is depicted by the
discussed in the Section 3.3.3 (Annual Results), this is due to the fact dotted line in Fig. 14.

P P
min ðTLðtiþ1 Þ  TLðti ÞÞ2 ¼ min ððDðtiþ1 Þ þ Xðtiþ1 ÞÞ  ðDðti Þ þ Xðti ÞÞÞ2
(i P P i
Dt  Xðti Þ ¼ B ¼ bn (14)
s:t: i n
Xðti Þ  Rðti Þ
 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597 595

For the second step, at each time or vehicle interval, the cost
function is modified by subtracting the target load from the cost will be used as the target load. Second, the same modified protocol
function defined in Section 3.2. (i.e., the costs shown in Figs. 8 and will be applied on the other target, the solution of ramp rate
9). This modification, see Equation (15), results in a cost function reducing.
that is essentially the gap between the current load and the target
load. The current load is updated by receiving and aggregating PEV 4.2.1. Valley filling as target
charging profiles sk(ti) in each step exactly the same way as before Fig. 16 shows the final load from the modified protocol with
while the target load is fixed. Fig. 14 also shows the initial modified 11:00 to 18:00 prioritized. Due to the prioritizing, the gap in the
cost functions for the two target loads when no PEV has arrived. small valleys is more negative than those of overnight valley,
steering early arriving PEVs to toward the smaller valleys. As these
TCk ðti Þ ¼ Ck ðti Þ  TLðti Þ are filled, or their different with the target becomes small, the
(15)
Ck ðti Þ ¼ Ck1 ðti Þ þ sk1 ðti Þ; C0 ðti Þ ¼ Dðti Þ updated cost function directs the PEVs to the overnight valley.
Generally, the modified cost function can be non-positive if the However, it ended up not filling the gap completely for the time
current load is less than the target level. The wider this gap in some slot 15:00 to 16:00. The reason is that the target load in the example
time slots, the more attractive it is for the PEVs to charge. By the is not feasible to achieve due to the small number of PEVs available.
analysis shown in the Appendix, PEVs will choose the lowest cost Note that the solution, in the dashed line, does not violate the
periods, which leads initially to late night time slots due to the maximum power constraint, but this constraint does not capture
largest gap between the two curves. As joining PEVs select these the actual challenge: some of the PEVs are already fully charged, so
time slots, the gap becomes small and the magnitude reaches the they are not able to take any more charge.
same as the ones in the afternoon, PEVs will then choose to charge
at the first gap. However, by then there might not be enough PEVs 4.2.2. Ramp rate reducing as target
available to fill the first gap, as the results show in Fig. 15. It is In Fig. 17, two 48-h net loads are shown with the left one the
exactly the same reason that the original protocol is not able to fill same as in the previous example and with the right one having
the afternoon valley in Fig. 13. Thus, an additional step of modifying multiple small valleys throughout the day. The big afternoon valley
the cost is needed to guarantee the afternoon gap is filled first. on the left pane implies substantial solar/wind generation during
For the third step, in order to prioritize the afternoon time slots, the day time while the two small valleys on the right pane indicate
G; grid operator can artificially scale the costs to be more attractive varying solar/wind generation. Compared to the example above,
than those overnight. Equation (16) shows the formulation of the only the target load is changed in order to have the minimum ramp
modified cost PCk(ti), that will be broadcast to PEVs, along with rate. The final loads with modified PEV protocol approached the
some requirements. targets very well. For the example on the left, the maximum dif-
ference from the target load is only 200 MW with only one
PCk ðti Þ ¼ TCk ðti ÞPðti Þ exception at the time slot from 15:00 to 16:00. That difference is
Pðti Þ ¼ 1cti ;  G; Pðti Þ > 1cti ˛G  (16)
around 400 MW and due to the same reason mentioned before that
PCk ðti Þ < PCk tj c TCk ðti Þ < 0; ti ˛G; tj ;G the target is not feasible to achieve. As for the example on the right,
PCk ðti Þ < PCk ðtiþ1 ÞcfTCk ðti Þ < 0; ti ˛G; tiþ1 ˛Gg this issue does not exist so the difference between final load and
target is examined to be smaller than 200 MW for all time slots.
The scaling vector, P(ti), has to be larger than one when ti be-
In summary, by slightly changing the cost function broadcast
longs to G; the time slots that are to be prioritized, which is 11:00 to
and prioritizing the time slots during the day time, the final load
18:00 in the example; it can be equal to one for the rest of the time
with PEVs can be controlled to approach a target load. To form an
slots. Second, from 11:00 to 18:00, the scaling values need to be
achievable target, grid operators need to estimate the overall
greater than the ratio between the depth of the overnight gap and
charging energy in (2), maximum charging power at each time slot
the depth of the afternoon gap, such that the afternoon gap is filled
in (4), and other possible constraints not included here. This
first. The scalar must be large enough that it produces lower costs in
modified protocol extended the capability of PEVs to change the
that window, even as the afternoon gap is partially filled. Third, for
grid load from valley filling to target load following.
the first gap, from 11:00 to 18:00, the same problem discussed
earlier can exist, that early PEVs will fill the deepest point, at
5. Discussion
around 15:00, and only then the rest of the gap. When the number
of PEVs is small, there is a possibility that sufficient PEVs are not
The results presented in this work often lead to Intermittent
available to fill the early part of the gap. P(ti), therefore, should
charging. As discussed in Refs. [11,32,33], there is little evidence
ensure the modified function PCk(ti) has a more negative value for
that intermittent charging is undesirable, indeed the avoidance of
the earlier time slots in the prioritized window G: In other words,
excess heating due to continuous charging over a long period, could
PCk(ti) has to have a positive slope in terms of ti if the target has not
be an important benefit. Nevertheless, if continuous charging has to
been reached. The last line in Equation (16) is the condition
be guaranteed, individual PEVs can first estimate the required
required to guarantee that the gap at the earlier time slots will be
charging time and then perform a line search to understand the
filled first by the PEVs. The precise values of P(ti) were manually
optimal starting point such that charging cost can be minimized.
tuned here, through a few trials, having the value on the order of
The rest of the protocol for cost updating will remain the same. For
10E6 for 11:00 and 10E3 for 18:00, and gradually decreasing
the overnight valley filling, this will show a very similar result
exponentially.
compared to intermittent charging, but have a small bump in the
middle of the valley. But if charging power is doubled to 6.6 kW, the
4.2. Results and analysis flat valley filling overnight can also be achieved.
The actual charging power for a specific PEV depends on its SOC.
In this section, two sets of final results will be shown. First, Generally, the maximum constant charging power is sustained
following the above example in Fig. 15, the solution of valley filling before the SOC reaches 80%e90% and lower charging power is
596 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597

applied until fully charged. In this study the algorithm to calculate 2. Using grid load directly as the cost function and updating it
individual charging profile does not consider the SOC dependent frequently enough, by either a fixed time interval or vehicle
charging profile, but assumes maximum power can be applied all amount, will lead to a flat final net load overnight for a relatively
the time. However, a simple change on the algorithm can accom- large time window. Updating the cost function every 30 min
plish different charging profiles. For continuous charging, PEVs can results in less than 300 MW variations on the final load during
estimate the SOC dependent charging profile, implement the line more than 7 h, for 90% of the days in a year. Also, the correlation
search to determine the charging starting point, and send back this of the aggregated charging loads from grid level valley filling
profile. For intermittent charging, PEVs can overestimate the and the proposed protocol is greater than 0.98.
charging energy, implement the linear optimization to determine 3. The computation and communication efforts required by the
charging profile with constant power, tailor the end of the charging proposed protocol are more modest than the existing works
profile as the battery specifies, and send back the final profile. For which require, potentially a large number of, iterations. Each
the grid operator, the protocol to aggregate PEV profiles and update PEV needs to compute its charge profile only once, performing a
the cost function does not change at all. simple linear optimization problem. It also needs to send the
Wind/solar power is subject to intermittency and difficult to charging profile back to the grid operator, where individual
predict [34e37]. If the forecast net load changes after most PEVs profiles are aggregated and loads are updated periodically.
have arrived and committed, it is unlikely that any simple protocol 4. Using the gap between the current load and final target load as
can ensure a flat final load. In this case, a possible solution is to the modified cost function and prioritizing the earlier time slots
reschedule a certain number of PEVs. The rescheduling would if necessary, a desired target net load can be approached similar
require an accurate forecast of the renewable power, coupled with to overnight valley filling.
some ranking and identifying PEVs with flexibility to be re-
scheduled. Such an approach would require significant extra
Acknowledgment
communication and is beyond the scope of this paper, which is
focused on simple protocols with minimal communication.
The authors would like to thank Brian Tarroja from University of
The issue of the actual cost of energy, charged to the customer,
California, Irvine, for providing data on grid demand and renewable
by the utility is also not investigated. If the protocol is focused on
power.
overnight charging, a low uniform pricing policy is likely, though
secondary valleys (reflecting intermittent renewables) may lead to
more complex pricing models. Another issue, not directly Appendix
addressed, is possible distribution constraints, particularly for the
residential transformer, given that the data used reflect statewide Here we present the main properties of the optimization algo-
numbers. The concept here can be used in relatively small networks rithm in (5)e(7). Each PEV will minimize its costs associated the
and grids, with little to no modifications. Finally, we do not consider following cost function.
vehicle-to-grid (V2G) operation. While theoretically, the optimi- X
zation can be modified to accommodate negative charging, such an Cðti Þxðti Þ
option would entail a variety of new constraints and conditions
(e.g., regarding grid stability, etc.) that are well beyond the scope of Subject to
this work, as well as near term possible technologies. 8P
< xðti Þ ¼ b
xðti Þ  0
:
6. Conclusions
xðti Þ  rðti Þ
The Lagrangian is
Realistic electric demand from CAISO for 2005 and solar and
X 
wind power under 30% renewable penetration assumptions were L ¼ CT x  n
T
xðti Þ  b  l x þ mT ðx  rÞ
used to generate the net load profile for the state of California. The
2009 National Household Travel Survey data and parameterized Applying the standard approach, KTT condition (which are both
PEVs operating and charging model were used to simulate the necessary and sufficient due to convexity)
charging demand and constraints with 2.1 million PEVs (10%
penetration). Optimization on the aggregated profile was formed in vL
¼ Cðti Þ  n  lðti Þ þ mðti Þ ¼ 0
terms of valley filling and ramp rate reducing. Optimization on vx
individual PEV charging was formed, as well, for the objective of
X 
minimizing charging cost. Two methods on cost function updating n xðti Þ  b ¼ 0
were performed along with results for both daily and annual basis.
The proposed protocol in this work shows a promising result in
terms of overnight valley filling and target load approaching. Based lðti Þxðti Þ ¼ 0
on the data, model and results above, the following conclusions can
be drawn: mðti Þðxðti Þ  rðti ÞÞ ¼ 0

1. At the demonstrated renewable penetration, the net load curve From

shows the biggest valley almost every night. The most effective
time for charging is the window from 20:00 to 7:00 in the next lðti Þxðti Þ ¼ 0/lðti Þ ¼ 0 or xðti Þ ¼ 0
day, corresponding to the availability of 80% of the total number Since we are only interested in charging time, we consider
of vehicles. With significant PEV market penetration (10%)
shown in this work, coordination between individual PEVs and xðti Þs0
the grid has to be made to avoid additional and prohibitively
expensive peak power periods. Then
 L. Zhang et al. / Journal of Power Sources 267 (2014) 584e597 597

References
lðti Þ ¼ 0
[1] SAE Hybrid e EV Committee, Hybrid Electric Vehicle (HEV) & Electric Vehicle
Then the KTT conditions become to (EV) Terminology, SAE, 2008.
[2] Mehrada Ehsani, Yimin Gao, Ali Emadi, Modern Electric, Hybrid Electric, and
8
i Þ  n þ mðti Þ ¼ 0
Fuel Cell Vehicles: Fundamentals, Theory, and Design, 2004.
< CðtP [3] Karel H. Jansen, Tim M. Brown, G. Scott Samuelsen, J. Power Sources 195
nð xðti Þ  bÞ ¼ 0
: (2010) 5409e5416.
mðti Þðxðti Þ  rðti ÞÞ ¼ 0 [4] Li Zhang, Tim Brown, G. Scott Samuelsen, J. Power Sources 196 (2011) 6559e
6566.
[5] E.D. Tate, Peter J. Savagian, The CO2 Benefits of Electrification E-REVs, PHEVs
If and Charging Scenarios, in: SAE World Congress & Exhibition, SAE, 2009.
[6] Stephanie Stockar, Pinak Tulpule, Vincenzo Marano, Giorgio Rizzoni, Energy,
mðti Þs0/xðti Þ ¼ rðti Þ Economical and Environmental Analysis of Plug-In Hybrids Electric Vehicles
Based on Common Driving Cycles, in: 9th International Conference on Engines
and Vehicles, SAE, 2009.
i.e., charging at maximum power. If [7] California Air Resources Board, Advanced Clean Cars Program, C.A.R. Board, 2012.
[8] M. Yilmaz, P. Krein, IEEE Trans. Power Electron 28 (2013) 2151e2169.
mðti Þ ¼ 0/xðti Þcan be different than rðti Þ but Cðti Þ ¼ n [9] Li Zhang, Tim Brown, Scott Samuelsen, J. Power Sources 240 (2013) 515e524.
[10] Joshua D. Eichman, Fabian Mueller, Brian Tarroja, Lori Smith Schell,
So the KTT condition shows either x(ti) ¼ r(ti) or C(ti) ¼ n. Scott Samuelsen, Energy (2013).
If C(ti)’s are distinct, then m(ti) ¼ 0 is possible for one time slot [11] Keenan Valentine, William G. Temple, K. Max Zhang, J. Power Sources 196
(2011) 10717e10726.
only, since there is only one n. [12] Changsun Ahn, Chiao-Ting Li, Huei Peng, J. Power Sources 196 (2011) 10369e
This shows that all other xðti Þ’s are at maximum value with 10379.
possible exception of 1. [13] Zhongjing Ma, Duncan Callaway, Ian Hiskens, Decentralized charging control
for large populations of plug-in electric vehicles, in: 49th IEEE Conference on
Claim: assuming distinct prices for each time slot, the algorithm Decision and Control (CDC), IEEE, 2010, pp. 206e212.
above picks the lowest cost time slots. Furthermore, the partial [14] Zhongjing Ma, Ian Hiskens, Duncan Callaway, A Decentralized MPC Strategy
time slot has the highest price among time slots used (but lower for Charging Large Populations of Plug-in Electric Vehicles, Preprints of the
18th IFAC World Congress Milano (Italy), August, 2011.
than those not used). [15] Lingwen Gan, U. Topcu, S. Low, Optimal decentralized protocol for electric
Proof: We start with the first part of the claim. We use the vehicle charging, in: 50th IEEE Conference on Decision and Control and Eu-
following notation for the charging and non-charging time slots, ropean Control Conference (CDC-ECC), 2011, pp. 5798e5804.
[16] Lingwen Gan, Ufuk Topcu, Steven H. Low, Stochastic distributed protocol for
respectively
electric vehicle charging with discrete charging rate, in: 2012 IEEE Power and
    Energy Society General Meeting, IEEE, 2012, pp. 1e8.
Ic ¼ jjx tj s0 [17] Alexander D. Hilshey, Pooya Rezaei, Paul DH. Hines, J. Frolik, Electric vehicle
charging: transformer impacts and smart, decentralized solutions, in: 2012
IEEE Power and Energy Society General Meeting, IEEE, 2012, pp. 1e8.
n   o
[18] Moein Moeini-Aghtaie, Ali Abbaspour, Mahmud Fotuhi-Firuzabad,
Inc ¼ j’ jx tj’ ¼ 0 Payman Dehghanian, PHEVs centralized/decentralized charging control
mechanisms: requirements and impacts, in: 2013, IEEE North American Po-
From the main property of the optimization in (6)e(8), for all wer Symposium (NAPS), 2013, pp. 1e6.
j˛Ic ; we have x(tj) ¼ r(tj) except at most one; i.e., maximum charge [19] M. González Vayá, Göran Andersson, Centralized and decentralized ap-
proaches to smart charging of plug-in vehicles, in: 2012 IEEE Power and
in all time slots with at most one partial charge. Energy Society General Meeting, IEEE, 2012, pp. 1e8.
Now assume there exists some j’ ˛Inc such that Cðtj’ Þ < Cðtk Þ for [20] Kazuhiro Kondo, Koichi Kojima, Mamoru Sasaki, Development of Charging
some k˛Ic ; i.e., one of the time slots with no charge has lower price System for Plug-in Hybrid Vehicles, EVS26, Los Angeles, California, 2012.
[21] Federal Highway Administration, The U.S. Department of Transportation,
than at least one of the charging times slots. Then consider the Introduction to the 2009 National Household Travel Survey (cited). Available
following: from: http://nhts.ornl.gov/introduction.shtml.
[22] U.S. Department of Energy, Chevrolet Volt Fuel Economy, 2012 (cited 2012).
  h   i Available from: http://www.fueleconomy.gov/feg/Find.do?
C tj’ ε þ Cðtk Þðxðtk Þ  εÞ ¼ C tj’  Cðtk Þ ε action¼sbs&id¼31618.
[23] CARB California Air Resource Board, EMFAC, 2007.
þ Cðtk Þxðtk Þ < Cðtk Þxðtk Þ [24] Paul Douglas, E. Stoltzfus, A. Gillette, J. Marks, California Public Utilities
Commission, 2009.
i.e., shifting the charging to the tj’ time slot reduces the cost, which [25] Paul Denholm, Michael Kuss, Robert M. Margolis, J. Power Sources 236 (2013)
350e356.
is not possible as it contradicts the optimality of the solution. This [26] Scott B. Peterson, Jay Apt, J.F. Whitacre, J. Power Sources 195 (2009) 2385e
shows the optimized solution picks the lowest cost time slots. 2392.
For the last part of the claim, we follow the same logic: suppose [27] Saeid Bashash, Scott J. Moura, Joel C. Forman, Hosam K. Fathy, J. Power Sources
196 (2010) 541e549.
tj was associated with partial charging, i.e., x(tj) < r(tj). Suppose [28] John Wang, Ping Liu, Jocelyn Hicks-Garner, Elena Sherman,
there exist tk such that C(tj) < C(tk) and x(tk) ¼ r(tk). Clearly, there Souren Soukiazian, Mark Verbrugge, Harshad Tataria, James Musser,
exists ε > 0 small enough such that Peter Finamore, J. Power Sources 196 (2011) 3942e3948.
[29] Robert C. Green Ii, Lingfeng Wang, Mansoor Alam, Renew. Sustain. Energy
     Rev. 15 (2011) 544e553.
0 < x tj þ ε  r tj [30] L. Fernandez Pieltain, T.G.S. Roman, R. Cossent, C.M. Domingo, P. Frias, IEEE
0  xðtk Þ  ε < rðtk Þ Trans. Power Syst. 26 (2011) 206e213.
[31] Michael Kuss, Tony Markel, William Kramer, 2010.
Similar to above, we note that [32] B.K. Purushothaman, U. Landau, J. Electrochem. Soc. 153 (2006) A533eA542.
[33] Jun Li, Edward Murphy, Jack Winnick, Paul A. Kohl, J. Power Sources 102
     (2001) 302e309.
C tj x tj þ ε þ Cðtk Þðxðtk Þ  εÞ [34] Ma Lei, Luan Shiyan, Jiang Chuanwen, Liu Hongling, Zhang Yan, Renew. Sus-
 
    tain. Energy Rev. 13 (2009) 915e920.
¼ C tj  Cðtk Þ ε þ C tj x tj þ Cðtk Þxðtk Þ [35] Ronan Doherty, Mark O’Malley, Quantifying reserve demands due to
    increasing wind power penetration, in: 2003 IEEE Bologna Power Tech Con-
< C tj x tj þ Cðtk Þxðtk Þ
ference Proceedings, vol. 2, IEEE, 2003, p. 5.
[36] Yuri V. Makarov, Pavel V. Etingov, Jian Ma, Zhenyu Huang,
which means shifting from tk to tj reduces the cost, which contra- Krishnappa Subbarao, IEEE Trans. Sustain. Energy 2 (2011) 433e442.
[37] Yuri V. Makarov, Clyde Loutan, Jian Ma, Phillip de Mello, IEEE Trans. Power
dicts optimality of the solution, implying that C(tj) > C(tk) for all
Syst. 24 (2009) 1039e1050.
fksjjk˛Ic g:.