Anda di halaman 1dari 6

2015 IEEE International Conference on Smart Grid Communications (SmartGridComm): Architectures, Control and Operation for

Smart Grids and Microgrids

Sizing and Control of Residential Solar Panel and


Battery
Neda Edalat and Mehul Motani Jean Walrand
Dept. of Electrical & Computer Engineering Dept. of Electrical Engineering & Computer Sciences
National University of Singapore University of California, Berkeley
Email: {neda, motani}@u.nus.edu Email: walrand@berkeley.edu

Abstract—Variable demand, dynamic electricity pricing and solar panel. This approach uses a small test panel to measure
intermittent renewable energy are sources of variability and the local availability of solar energy as a function of time and
uncertainty in smart grid systems. Energy storage devices can be the house smart meter readings to track the actual load. These
used to absorb these fluctuations and reduce the overall cost. The
goal is to find the best policy for storing, buying, selling and using measurements are made for some time prior to making the
energy in the presence of this variability and uncertainty. This investment decisions. An algorithm then simulates the system
paper explores a comprehensive methodology to study the effect for different values of policy parameters and sizes of the
of integrating renewable sources and energy storage in reducing battery and panel and evaluates the corresponding operating
the cost of smart grid operation and the sizing and control of and capital costs. The outcome of these measurement-driven
such systems. In this paper, we consider grid-connected dwellings
with a solar panel and battery as an example scenario. Numerical simulations is a choice of policy parameters and size of the
results illustrate the performance of our proposed methodology. battery and solar panel. Note that after the solar activity has
been collected, the measurements can be combined with the
I. I NTRODUCTION smart meter readings of other houses in a similar location to
select their appropriate system.
Uncertainties associated with renewable sources create a
We also analyze how well the PP performs by comparing
challenge for their integration into the grid [1]. Energy storage
them to the solution of a Markov decision problem that
can play an important role by mitigating the variable load,
assumes a genie who reveals hidden state information.
renewable source and electricity prices [2]-[5].
The goal of the proposed policies is to minimize the long-
The analysis of suitable policies for storing, buying, sell-
term average cost while fully meeting the variable demand.
ing and using energy in the presence of many sources of
The cost includes the capital cost of solar panels and batteries,
uncertainty is complex. Our main objective is to develop a
as well as the operating cost of buying energy from the grid.
simple approach that enables the study of various scenarios
for integrating energy storage and renewable energy sources A. Related Work
in smart grid systems.
Researchers have studied approaches for storage manage-
In this paper, we focus on grid-connected houses equipped
ment and control in order to optimize several features in
with a solar panel and a battery. The battery is useful for:
smart grid systems. They have explored using energy storage
• Reducing the variability of the demand from the grid. for minimizing energy imbalance (the differences between
This variability is undesirable for the utility company that demand and local energy). The authors of [7] proposed an
may impose a tariff that penalizes peak demand. approach to reduce reserve energy requirement in power
• Adjusting for the time-of-day variations of the renewable system dispatch. They formulate the power imbalance problem
source and load that are typically out of phase. for each timescale as an infinite horizon stochastic control
• Exploiting time-of-day pricing by buying electricity when problem and show that a greedy policy minimizes the average
it is cheap. magnitude of the residual power imbalance. Their analysis
Classically, energy storage for solar panels is mostly designed applies to the case in which the utility company owns and op-
for stand-alone applications. This means that an amount of erates the storage. They use the prediction for multi time scale
energy must be stored to fill at least the production gap of of renewable generation. The optimal power flow problem with
several cloudy weather days. If the dwelling is connected to energy storage is formulated in [8] assuming deterministic load
the grid, an energy storage system should not necessarily cope with the objective of minimizing the total quadratic generation
with a long period of low solar panel production, as the grid cost. [9] addresses algorithms for minimizing the long-term
is available as a back-up. Hence, the storage and solar panel energy cost in the presence of large storages, using online
sizing criteria are different from autonomous systems. dynamic control approach. They assume that the prices and
First, we develop a methodology for selecting a good demands follow a fixed but unknown distribution. Moreover,
parametric policy (PP) for storing, buying, selling and using they assume there are no renewable sources and the battery
energy and the corresponding optimum size of the battery and used only for the purpose of arbitrage. Similarly, in [10]

978-1-4673-8289-2/15/$31.00 ©2015 IEEE 527


2015 IEEE International Conference on Smart Grid Communications (SmartGridComm): Architectures, Control and Operation for
Smart Grids and Microgrids

the authors consider the problem of energy storage from the


point of view of the grid operator, and propose a threshold
policy that is shown to be asymptotically optimal for the large
size of the storage unit. In [11], the algorithm leverages a
prediction model they develop, which forecasts future demand
using statistical machine learning techniques given the next- Fig. 1: System model.
day prices. There are some other relevant works such as
[12] for storage with demand response and [13] for operating
storage co-located with end-user demands which is a similar At time slot n, the difference between the available energy
approach used in data centers. The authors of [17] study the Gn + Rn S and the demand Ln is stored in the battery. Thus,
tradeoff between the use of reserves and the energy loss in a Wn+1 = [Wn + Gn + Rn S − Ln ]B
0
power grid and develop a scheduling algorithm for small or
moderate storage. where [x]B0 := max{0, min{x, B}}.
Controlling the residential energy storage has been proposed The main objective is to explore approaches for choosing
in [14] - [16]. The authors analyze the benefits of buying Gn , the size of the battery B and the size of the renewable
energy when it is cheap and selling it later to the grid for higher source S such that the overall cost is minimized. The cost
price, assuming that the prices and statistics are known in a contains the capital cost of solar panel and energy storage
finite horizon. They formulate the deterministic optimization devices as well as the operating cost of buying energy from
problem and solve it by linear programing methods. the grid.
The above-mentioned algorithms are mainly designed for III. P ROBLEM F ORMULATION
large battery capacity and do not take into account the cost of Considering this model, one can precisely formulate the
both battery and solar panel. To the best of our knowledge, this advantage of using the battery for grid-connected systems
is the first work that explores a comprehensive methodology to equipped with solar panels. Assuming that we want to meet
study the effect of using energy storage devices and renewable the load Ln all the time, the problem is to minimize the overall
energy on the overall cost reduction of operating the smart grid average capital and operating costs. This leads to the problem
systems. The output of proposed policy and analysis may be statement described below.
used for making the investment decisions. N
This paper is structured as follows. We introduce the system 1 
Minimize γS + βB + lim E(cn (Gn )) (1)
model and problem formulation in Sections II and III. In N →∞ N
n=1
Section IV, we present the proposed PP for the problem over S, B, Gn
formulation. Numerical results for PP and DP are provided in
subject to Gn + Rn S + Wn ≥ Ln , ∀n ≥ 1
Section V. Section VI concludes and summarizes this paper.
where cn (.) denotes the cost of buying energy from the grid
II. S YSTEM M ODEL and it is strictly monotonic increasing on the amount of
energy and periodic as a function of time n. Here, γ and
We consider a grid-connected system with a renewable β are the amortized costs per time slot of a unit-size solar
source and a battery. The discrete time system model is panel and battery, respectively. The control policy is of the
sketched in Figure 1. The time unit of this model is one form Gn = fn (Rn , W n , Ln ), ∀n ≥ 1. In this expression,
hour and a time slot is denoted by n ∈ {1, 2, . . .}. In this Rn = {R1 , R2 , . . . , Rn } and similarly for W n and Ln .
model, solar energy is used as a renewable energy source. We study the effect of using energy storage in the different
S denotes the size of the solar panel and Rn is the energy scenarios:
that a unit-size solar panel produces. Hence, Rn S is the Scenario 1: In this scenario, the amount of the energy that
total amount of energy produced in time slot n. The energy one buys is independent of the time-of-day pricing. The goal is
bought from the grid at time n is Gn . The energy consumed to regulate the demand from the grid. In order to motivate the
at time n is denoted as Ln . The methodology does not user to smooth out the demand, one uses a convex increasing
assume that a model is known for the random sequences cost of energy such as cn (Gn ) = G2n .
{Rn , n ≥ 1} and {Ln , n ≥ 1}. We evaluate the performance Scenario 2: In this scenario, the cost function accounts for
of the methodology when these sequences are functions of the time-of-day pricing of electricity. The goal is to adjust
irreducible time-homogeneous Markov chains X and Y on for time-of-day variation of the price, renewable source and
some finite state space X and Y, respectively. This evaluation load. In this case, the cost of energy is a linear function of the
tells us by how much one could further reduce the cost if one amount of the energy that one buys that depends periodically
had a more precise stochastic model of the weather, demand, on time n, such as cn (Gn ) = an Gn where an is periodic.
and prices. It turns out that this additional reduction is small, Scenario 3: In this scenario, one includes the possibility of
so that our oblivious methodology is effective. selling back electricity to the grid. We assume that the selling
At time, n ≥ 1, the battery with capacity B has accumulated price is κE(n) during time slot n, where E(n) is the buying
an amount of energy denoted as Wn ∈ {0, 1, . . . , B}. price and κ ∈ [0, 1].

528
2015 IEEE International Conference on Smart Grid Communications (SmartGridComm): Architectures, Control and Operation for
Smart Grids and Microgrids

IV. A PPROXIMATION λ(n) = λ1 if electricity is cheap during time slot n and let
One might formate the problem as a Markov decision λ(n) = λ2 otherwise. The policy buys Gn from the grid.
problem with incomplete observations that would then give the Define
optimal control policy Gn . This approach is not satisfactory An = W n + R n S − L n .
for a number of reasons:
Then, if An < 0,
• The optimal policy is complicated and depends on un-
known parameters. Since, the control policy Gn at each Gn = max{λ(n), −An }, Wn+1 = max{0, λ(n) − An }.
time is a function of (Rn , W n , Ln ), solving the DP
equations requires discretizing these distributions, which Also, if An > 0,
results in a very large state space.
• Lack of guidelines and insight. This approach is brute-
Gn = min{λ(n), max{B−An , 0}}, Bn+1 = min{B, An +λ(n)}.
force and provides numerical results with no insight into
For instance, if An > 0, then the renewable energy plus the
how they depend on various aspects of the system.
energy in the battery suffice to serve the load. In that case,
For these various reasons, our goal is to develop a more the battery charge reaches the value max{B, An } without the
insightful and scalable methodology that may suggest practical grid. One then buys max{λ(n), B − max{B, An }} to further
learning mechanisms. recharge the battery. The other cases are similar.
The cost of the system at time n is then
A. Parametric Policies (PP)
In this section, we discuss a measurement-based approach Cn = E(n)Gn + γS + βB
to optimize the design and control of the system. First, we
collect measurement traces for the specific dwelling, or we use recall that E(n) is the unit price of the electricity during time
measurements previously made in the same neighborhood that slot n, γ is the amortized cost of a unit-size solar panel and β
we recalibrate. We describe the former case. One first installs that of unit-size battery per time slot. We assume for simplicity
a small test solar panel on the roof of the house to record that the costs are linear in the size, but that assumption can
a time series of the solar power that such a panel produces. be relaxed easily to conform to actual prices.
In parallel, one uses the smart meter readings to record the For the third scenario, we include the possibility of selling
electrical consumption of the house over time. This monitoring back electricity to the grid. We assume that the selling price
produces two time series: Rn and Ln . is κE(n) during time slot n, where E(n) is the buying price
Given the traces (Rn , Ln ) ∀n ≥ 1, one performs parallel and κ ∈ [0, 1]. If κ = 0, then the optimal policy will not
trace-driven simulations for different solar panel and battery sell electricity. The policy sells electricity if the battery would
sizes and parameters of control policies. These simulations overflow otherwise and also if there is some available energy
track the operating costs of these different systems and allow and if λ(n) < 0. The policy buys G(n) from the grid.
to identify the values that result in the smallest cost. ⎧

⎪ min{B − An , λ(n)}, if An > B or
To implement this methodology, one must define parametric ⎪

⎨ 0 < An < B and λ(n) > 0;
control policies. For the first scenario (as described in Section
Gn = max{−An , λ(n)}, if 0 < An < B and
III), the strategy for different values of α, S and B is as ⎪


⎪ λ(n) < 0;
follows: we buy electricity at constant rate Gn = α, ∀n ≥ 1 ⎩
−An , if An < 0.
except that we stop buying when the battery is full and we
buy Ln − Rn S when the battery is empty. We assume that the Also,
cost of buying electricity from the grid is a non-linear function Wn+1 = An + Gn
of the amount of energy. These simulations keep track of the
average cost of buying electricity for different values of B and and the cost of operating the system during time slot n is Cn
S. From these simulations, one can determine the best values where
of those parameters. This strategy corresponds to a simple
Cn = En Gn 1{Gn > 0} + κEn Gn 1{Gn < 0}
policy for buying energy from the grid.
+γS + βB.
For the second scenario, the PP is slightly different. Es-
sentially, the policy buys electricity at rate λ1 and λ2 during From the trace-driven simulation, one can then compute an
the night (off-peak) and day (on-peak), respectively. However, average cost per hour C(S, B, λ1 , λ2 ) as follows:
this policy is adjusted to meet the load and not overflow the
N −1
battery. 1 
More precisely, the policy is defined as follows. Let Wn be C(S, B, λ1 , λ2 ) = Cn .
N n=0
the charge of the battery at time n and B the capacity of the
battery. Assume that a solar panel of size S would produce Using these results, one can determine the values of
the power Rn S, where Rn is the power that the test (unit- (B, S, λ1 , λ2 ) that minimize the capital and operating costs
size) panel produces. Finally, recall that the load is Ln . Let of the system.

529
2015 IEEE International Conference on Smart Grid Communications (SmartGridComm): Architectures, Control and Operation for
Smart Grids and Microgrids

Item Cost Units Typical Values


Electricity, night κ1 ¢/kWh 20
Electricity, daytime κ2 ¢/kWh 30
Battery β ¢/kWh 7
Average Solar Panel γ ¢/kWh 15
Average Load l kWh/day 30

TABLE I: Parameters of the system.

Fig. 2: The model used for the simulation.


V. N UMERICAL R ESULTS
We now compare numerical results for both the PP and DP.
The traces (Rn , Ln ) are generated as follows. There are K so-
lar power and K load profiles, which indicate the solar power
and load, resp., during the 24 hours of the day.. The kth solar
power profile and load profile are (R(k, 1), . . . , R(k, 24)) and
(L(k, 1), . . . , L(k, 24)), resp. The solar power profile and load
profile for day d is X(d) and Y (d), resp. The sequences X(d)
and Y (d) are independent Markov chains with given transition (a) Daily load. (b) Daily renewable energy.
matrices. Recall that these models are used to evaluate the
policies, but that the policies are oblivious to these models. Fig. 3: Measurement traces.
Figure 2 shows the model (in which K = 3) used for the
simulation. This model is used to generate traces that have
certain features. For instance, there is typically a time offset Bellman’s equation for the long-term average cost is [22]:
between the peak solar power and peak load. Similarly, there 
c∗ + H(x) = max[c(x, u) + P (x, x ; u)H(x )]. (2)
is a time offset between cheap electricity prices electricity and u∈U
x
peak consumption. Note that these profiles are used to generate
traces, but they are fairly arbitrary. The representative traces In this expression, c∗ is the optimum cost, H(x) is the
for the renewable energy and load are shown in Figure 3. differential cost starting from state x and c(x, u) is the cost for
The key parameters of the system are the various costs, taking action u in state x. We use value iteration for average
shown in Table I. The typical values are obtained as follows. cost to compute the optimal policy.
κ1 and κ2 are from [18]. For the battery cost, we consulted The system states for the first scenario introduced in Section
catalogs and found that 12 Volt, 100 Ah battery costs about III are (Wn , Rn , Ln ). At each time slot n, the action is
$150. This means $150 for 1.2 kWh, or $120 per kWh. We made based on the states at that time. The action denoted
assume that this battery will last 1800 days, which gives a as Un indicates units of the energy drained from the battery
cost of 7 cents per kWh per day. Let’s consider a simple and the one dispatched from the grid to satisfy the load.
example with a battery and no solar panel to clarify this cost. This simulation runs for different values of B and S. Then,
With a small battery capacity of 1 kWh, one may buy 1 kWh the optimum action, B ∗ and S ∗ are calculated based on the

electricity during a night and use it later during a day. Then, minimum average total cost CDP 1
.
κ1 − κ2 − β cents can be saved per day. For solar panels, The system states are slightly different for the second and
we use [20] that quotes a net cost of $10, 000 for a system third scenarios in Section III. The periodic time is also a part
that produces about 4, 000 kWh per year. We assume that the of states as (Tn , Wn , Rn , Ln ) where Tn = n mod 24. This is
system is paid at the rate of $600 per year, over 25 years, due to the time-of-day pricing in the model. The action is the
which gives an amortized cost equal to 15 cents per kWh per same as before. The average total costs for these scenarios are
day, including rebates. For the average load, we used [21] that denoted as CDP2 and CDP3 , respectively. The parameters of
quotes an average household use of 903 kWh per month, or 30 the system are from Table I. We initially consider the transition
kWh per day. For the sake of performance evaluation, different probability as a = b = c = d = e = f = 0.3, a = b = c =
ranges of these values might be used. d = e = f  = 0.4.
For the first scenario, Figure 4 shows the effect of the battery
A. Dynamic Programming (DP) size B on the different costs when S = 3. By increasing B,
To verify the effectiveness of the simple PP, we compare the cost of buying from the grid decreases and the battery cost
them to more complex policies derived using DP. These increases. The optimum value is the battery size that results
complex policies assume complete knowledge of the states in minimum total cost. Figure 5 shows the effect of S and
X(d) and Y (d), and the transition matrices and sequences B on the average total cost (in ¢/kWh). It can be seen that
R(d, n) and L(d, n). Thus, these complex policies represent by increasing the battery size, the optimum value of the solar
the best that one could do will full information. If our policies panel size S ∗ slightly increases. In other words, by harvesting
perform well compared to these ideal policies, we can be more energy, larger battery sizes may result in lower costs.
reassured that not much is lost because of the simplicity. Considering the first scenario, it is observed in Table II that

530
2015 IEEE International Conference on Smart Grid Communications (SmartGridComm): Architectures, Control and Operation for
Smart Grids and Microgrids

a d B∗ ∗
CDP 1
(¢/kWh)
0.3 0.3 8 19.1
0.5 0.6 10 20.2
0.6 0.7 12 22.4
0.6 0.8 12 23.6

TABLE II: Different weather conditions.

B∗ S∗ α∗ C1∗ (cent/kWh)
β=5 8 1 4 26.1
β=7 7 2 3 28.5
β=9 5 2 2 30.4
β = 12 4 2 2 31.7

TABLE III: Decision variables vs. β (¢/kWh) for Scenario 1

Fig. 4: Example: Cost (¢/kWh) vs. B when S = 3. β, the battery size decreases and the size of the solar panel is
almost the same. Moreover, α decreases and the average cost
increases as β increases. This is because, the smaller battery
can store a lower value of the constant electricity rate α.
The result for scenario 2 is presented in Table IV. The
decision variables for this problem are λ1 , λ2 , S and B. C2∗
denotes the optimum cost for this scenario. We set κ1 = 20
and evaluate the performance for varying (κ2 −κ1 ). Increasing
(κ2 −κ1 ) leads to larger battery sizes. This is because one may
be able to store more energy during the night when electricity
is comparably cheaper. Similarly, the value for λ1 increases
(and λ2 decreases) as (κ2 − κ1 ) increases.
Figure 7 shows the amount that one can save by having the
opportunity to sell back to the grid. Recall that one may sell
back at the cost κE(n), where E(n) is the rate for buying
Fig. 5: Example: Average total cost (¢/kWh) vs. S and B. electricity at time n. For simplicity, E(n) is κ1 and κ2 for
the electricity price at night and day, respectively. The costs
are defined in Table I. The total cost decreases as κ increases,
if the probability of going to the cloudy state increases (by because the selling back rate increases. Hence, one may use
increasing a and d), the optimum battery size and average larger batteries to store more than the daily load to sell it back
∗ to the grid. Given these parameters of the system, the battery
total cost (CDP 1
) increase. This is because there is a higher
chance for larger battery sizes to store more during the sunny cannot be larger than a certain value.
state and save it for the cloudy state. However, by increasing Figure 8 shows that, as β varies, the costs of the PP are very
a and d, B ∗ may not increase from the certain level, because
there is insufficient renewable energy to fill up the battery.
Hence, one may not gain by using larger battery sizes.
Using the first and second scenarios, the average total costs
∗ ∗
from DP (CDP 1
and CDP 2
) are compared with the no battery
∗ ∗
and no solar panel case (CN O1 and CN O2 ) (in Figure 6).
This shows how much one can earn by using solar panels and
batteries with different costs. The reason for the convergence
of the average total cost is that by increasing β, the battery
size approaches zero. Considering the effect of solar panel,
the costs from DP are always lower than the costs from the
no battery and no solar panel case.

B. Results for Parametric Policies


We now present results for the the PP. The decision variables
for the scenario 1 are α, S and B. Table III shows these values
and the average total cost versus β (¢/kWh). C1∗ denotes the Fig. 6: Comparison between different costs over β.
optimum cost for this problem. Not surprisingly, by increasing

531
2015 IEEE International Conference on Smart Grid Communications (SmartGridComm): Architectures, Control and Operation for
Smart Grids and Microgrids

B∗ S∗ λ∗1 λ∗2 C2∗ (¢/kWh)


(κ2 − κ1 )=5 8 2 3 2 17.1
(κ2 − κ1 )=8 10 3 4 1 19.3
(κ2 − κ1 )=10 12 3 4 0 20.4
(κ2 − κ1 )=12 13 3 5 0 23.6

TABLE IV: Decision variables vs. (κ2 − κ1 ) for scenario 2.

Fig. 8: Comparison of the costs from PP and DP vs. β.

[3] G. Barbose and C. Goldman, “A survey of utility experience with real time
pricing.”Berkeley National Laboratory, Tech. Rep. LBNL-54238, 2004.
[4] H. Allcott, “Rethinking real-time electricity pricing.” Resource and En-
ergy Economics 33.4 (2011): 820-842.
Fig. 7: Cost reduction by selling back to the grid. [5] A-H. Mohsenian-Rad et al., “Autonomous demand-side management
based on game-theoretic energy consumption scheduling for the future
smart grid.” IEEE Trans. Smart Grid, no. 3 (2010): 320-331.
[6] W. Ketter et al., “The 2013 Power Trading Agent Competition.” ERIM
near to the optimum costs from DP. However, the gap between Report Series Reference No. ERS-2013-006-LIS (2013).
[7] H. I. Su and A. El Gamal, “Modeling and Analysis of the Role of Energy
these costs increases as the cost of the battery increases. This Storage for Renewable Integration: Power Balancing.” IEEE Trans. Power
shows that the PP may not be good for smaller battery sizes Syst., 28(4):4109-4117, 2013.
where more accurate battery energy management is required. [8] D. Gayme and U. Topcu, “Optimal power flow with distributed energy
storage dynamics.” In Proc. of Amer. Control Conf., 2011, pp. 1536-1542.
Moreover, it can be seen that with lower battery cost one can [9] P.M. van de Ven et al., “Optimal control of end-user energy storage.”
save more by selling electricity back to the grid. Hence, the IEEE Trans. Smart Grid, vol. 4, no. 2 (2013): 789-797.
cost gap between scenarios 2 and 3 decreases as β increases. [10] I. Koutsopoulos et al., “Optimal energy storage control policies for the
smart power grid.” In Proc. IEEE Int. Conf. on Smart Grid Comm., pp.
VI. C ONCLUSION 475-480. IEEE, 2011.
[11] A. Mishra et al., “SmartCharge: cutting the electricity bill in smart homes
In this paper, we develop a methodology for selecting good with energy storage.” In Proc. of ACM 3rd Int. Conf. on Future Energy
PP for storing, buying, selling and using energy and the Sys.: Where Energy, Comp. and Comm. Meet, p. 29., 2012.
[12] R. Urgaonkar et al., “Optimal Power Cost Management Using Stored
corresponding optimum sizes of the battery and solar panel for Energy in Data Centers.”, ACM SIGMETRICS 2011, pp. 221-232, 2011.
grid-connected households. We formulate a Markov decision [13] L. Huang, J. Walrand, and K. Ramchandran, “Optimal Demand Re-
problem, that is inherently complex, for minimizing the overall sponse with Energy Storage Management.” In Proc. IEEE Int. Conf. on
Smart Grid Comm., pp. 61-66. IEEE, 2012.
long-term average cost and show that simple PP lead to [14] H. Saele and O.S. Grande, “Demand response from household cus-
promising solutions. Numerical results show that the the PP tomers: experiences from a pilot study in Norway.” IEEE Trans. Smart
perform well compared to the results from DP. This model can Grid, 2.1 (2011): 102-109.
[15] K. Ahlert and C. Van Dinther, “Sensitivity analysis of the economic
be extended to be able to analyze the battery degradation cost benefits from electricity storage at the end consumer level.” In PowerTech,
by adding a suitable term to the capital cost in the problem 2009 IEEE Bucharest, pp. 1-8. IEEE, 2009.
formulation. [16] J. Rajasekharan, and V. Koivunen, “Optimal Energy Consumption
Model for Smart Grid Households with Energy Storage.” arXiv preprint
arXiv:1310.3424 (2013).
ACKNOWLEDGMENT [17] N. Gast, D.-C. Tomozei and J.-Y. Le Boudec. “Optimal Generation and
The authors would like to thank Longbo Huang for suggest- Storage Scheduling in the Presence of Renewable Forecast Uncertainties.”
In IEEE Trans. Smart Grid, vol. 5, num. 3, p. 1328-1339, 2014.
ing the use of weather and demand scenarios for evaluating the [18] PG&E Residential Time of Use Pricing. http://goo.gl/dGVnAf
policies. This research was supported in part by the National [19] W IKIPEDIA Electric Vehicle battery. http://goo.gl/6cPmGN
University of Singapore and by MURI grant BAA 07-036.18. [20] E NERGY I NFORMATIVE How Much Do Solar Panels Cost. http://goo.
gl/B2Fc2K
R EFERENCES [21] U.S. E NERGY I NFORMATION A DMINISTRATION How Much Electricity
Does an American Home Use? http://goo.gl/R7ycMj
[1] Y. Yan et al., “A survey on smart grid communication infrastructures: [22] D. P. Bertsekas, “Dynamic Programming and Optimal Control.” Athena
Motivations, requirements and challenges.” IEEE Commun. Surveys Scientific, vol. 2, 3rd Edition, 2007.
Tuts., no. 1 (2013): 5-20.
[2] M.H. Albadi and E.F. El-Saadany, “A summary of demand response in
electricity markets.” Elec. Power Sys. Res., no. 11 (2008): 1989-1996.

532