Optimal Real-time Pricing Algorithm Based on
Utility Maximization for Smart Grid
Pedram Samadi, Amir-Hamed Mohsenian-Rad, Robert Schober, Vincent W.S. Wong, and Juri Jatskevich
Department of Electrical and Computer Engineering
The University of British Columbia, Vancouver, Canada
E-mail:{psamadi, hamed, rschober, vincentw, jurij}@ece.ubc.ca
Abstract—In this paper, we consider a smart power infras-
tructure, where several subscribers share a common energy
source. Each subscriber is equipped with an energy consumption
controller (ECC) unit as part of its smart meter. Each smart meter
is connected to not only the power grid but also a communication
infrastructure such as a local area network. This allows two-way
communication among smart meters. Considering the importance
of energy pricing as an essential tool to develop efficient demand
side management strategies, we propose a novel real-time pricing
algorithm for the future smart grid. We focus on the interactions
between the smart meters and the energy provider through
the exchange of control messages which contain subscribers’
energy consumption and the real-time price information. First,
we analytically model the subscribers’ preferences and their
energy consumption patterns in form of carefully selected utility
functions based on concepts from microeconomics. Second, we
propose a distributed algorithm which automatically manages
the interactions among the ECC units at the smart meters and
the energy provider. The algorithm finds the optimal energy
consumption levels for each subscriber to maximize the aggregate
utility of all subscribers in the system in a fair and efficient
fashion. Finally, we show that the energy provider can encourage
some desirable consumption patterns among the subscribers by
means of the proposed real-time pricing interactions. Simulation
results confirm that the proposed distributed algorithm can
potentially benefit both subscribers and the energy provider.
I. INTRODUCTION
Electricity is currently provided through an infrastructure
consisting of utility companies, power plants, and transmission
lines which serve millions of customers. For example, the
electric power grid in the United States includes more than
3,100 electric utilities operating more than 10,000 power
plants, and there are about 157,000 miles of high voltage
electric transmission lines which bring energy to more than
131 million customers [1]. The dependency of almost all parts
of industry and different aspects of our life on electrical energy
makes this massive infrastructure a strategic entity.
Given the increased expectations of customers, both in
quality and quantity [1], the limited energy resources, and the
lengthy and expensive process of exploiting new resources, the
reliability of the grid has been put in danger and there is a
need to develop new methods to increase the grid efficiency.
Currently, the electricity consumption is not efficient in most
buildings (e.g., due to poor thermal isolation). This results in
the waste of a l arge amount of natural resources, since most of
the electricity consumption occurs in buildings [2]. In addition,
the arising of new types of demand such as plug-in hybrid
electric vehicles (PHEVs), which can potentially double the
average household load, have further increased the need to
develop new methods for demand side management (DSM).
There is a wide range of DSM techniques such as voluntary
load management programs [3]–[5] and direct load control [6].
However, smart pricing is known as one of the most common
tools that can encourage users to consume wisely and more
efficiently. Given the recent increases in the price of energy,
the users are more willing to improve the insulation conditions
of their buildings or try to shift the energy consumption
schedule of their high-load household appliances to off-peak
hours. DSM has been considered since the early 1980s [7]–
[11]. DSM can be used as a tool for load shaping, where the
electricity demand is being re-distributed over a certain period
of time (e.g., time-of-day, day-of-week). Broad categories of
load shaping objectives include peak clipping, load shifting,
valley filling, strategic conservation, and flexible load shaping
[7]. For example, peak clipping includes direct load control of
the utilities on customers’ appliances to reduce the peak load.
Several pricing schemes have already been proposed in the
smart grid literature. In general, flat pricing, peak load pricing,
and adaptive pricing are among the most popular approaches
to pricing which have been practiced extensively [12]–[15].
Flat pricing refers to those methods where the utility company
announces a fixed price for all periods. In peak load pricing,
the intended cycle is divided into several periods and a distinct
price value for each period is announced at the beginning of
the operation [14]. On the other hand, in adaptive pricing,
instead of announcing a pre-determined price for each period
of operation at the beginning of the day, the exact price value
for each period is calculated in real-time and is announced
only at the beginning of each operation period. Clearly, in
this method, the realization of random events and the reaction
of users with respect to the previous prices will influence the
price in the upcoming operation periods [12].
Based on a report of the U.S. Department of Energy
[16], smart grid is an electricity delivery system enhanced
with communication facilities and information technologies
to enable more efficient and reliable grid operation with an
improved customer service and a cleaner environment. By
exploiting the two-way communication capabilities of smart
meters it becomes possible to replace the current power system
with a more intelligent infrastructure [17]. From this and given
the importance of demand side management, in this paper,
we focus on the real-time interactions among subscribers and
the energy provider and introduce a novel r eal-time pricing
algorithm for the future smart grid. The contributions of this
paper can be summarized as follows:
• We propose a real-time pricing algorithm for DSM pro-
grams to encourage desired energy consumption behav-
iors among users and to keep the total consumption level
below the power generation capacity.
• In our system model, the subscribers and the energy
provider automatically interact with each other through a
limited number of message exchanges and by running a
distributed algorithm to find the optimal energy consump-
tion level for each subscriber, the optimal price values to
be advertised by the energy provider, and also the optimal
generating capacity for the energy provider.
• We model the subscriber’s preferences and their energy
consumption patterns in form of carefully selected utility
functions based on concepts from microeconomics.
• We formulate the real-time pricing as an optimization
problem to maximize the aggregate utility of all sub-
scribers in the system while minimizing the imposed
energy cost to the energy provider. Moreover, we include
constraints to limit the total energy consumption level of
all users to the total electricity generation capacity of the
system offered by the energy provider.
• We prove the existence and the uniqueness of the optimal
solution for the formulated optimization problem.
• Simulation results confirm that both subscribers and the
energy provider will benefit from the proposed algorithm.
This paper is organized as follows. The system model is
presented in Section II. In Section III, we formulate our design
as a convex optimization problem and propose a distributed
pricing algorithm. Simulation results are given in Section IV,
and conclusions are drawn in Section V.
II. S
YSTEM MODEL
Consider a smart power system consisting of a single energy
provider, several load subscribers or users, and a regulatory
authority. For each user, we assume that there is an energy
consumption controller (ECC) unit which is embedded in the
user’s smart meter. The role of the ECC is to control the user’s
power consumption, and to coordinate each user with other
users and also with the energy provider. All ECC units are
connected to each other and to the energy provider through a
communication infrastructure such as a local area network.
The intended time cycle for the operation of the users is
divided into K time slots, where K |K|, and K is the set
of all time slots. This division can be based on the behavior
of the users and their power demand pattern: peak load time
slots, valley load time slots, and normal load time slots. Also,
let N denote the set of all users, where N |N |. For each
user i ∈N,letx
k
i
denote the amount of power consumed by
user i in time slot k. For each subscriber i ∈N and each time
slot k ∈K, we define the power consumption interval I
k
i
as
I
k
i
[m
k
i
,M
k
i
] (1)
and the consumed power x
k
i
has to satisfy m
k
i
≤ x
k
i
≤ M
k
i
.
Here, m
k
i
and M
k
i
denote the minimum and the maximum
power consumption of user i, respectively. The minimum
power consumption level may represent the load from ap-
pliances such as refrigerator which always need to be on
during the day. The maximum power consumption level may
also represent the total power consumption level of household
appliances assuming that all appliances are on.
The regulatory authority ensures that the energy provider
will provide the minimum capacity to cover the minimum
power requirements of all users L
min
k
in each time slot.
L
min
k
i∈N
m
k
i
, ∀k ∈K. (2)
The generation capacity in each time slot k ∈Kis denoted by
L
k
, which may differ among time slots. We also define L
max
k
as the maximum generating capacity in each time slot k ∈K.
A. User Preference and Utility Function
Each individual subscriber in a power system is an en-
tity which can behave independently. The energy demand
of each subscriber may vary based on different parameters.
For example, we can take into account the time of day,
climate conditions, and also the price of electricity. The energy
demand also depends on the type of the users. For example,
household users may have different responses to the same
price than industrial users. The different response of different
users to various price scenarios can be modeled analytically by
adopting the concept of utility function from microeconomics
[18]. In fact, we can model the behavior of different users
through their different choices of utility functions [4]. For
all users, we represent the corresponding utility function as
U(x, ω), where x is the power consumption level of the user
and ω is a parameter which may vary among users and also
at different times of the day. More formally, for each user, the
utility function represents the level of satisfaction obtained by
the user as a function of its power consumption. We assume
that the utility functions fulfill the following properties:
1) Property I: Utility functions are non-decreasing. That
is, users are always interested to consume more power if
possible until they reach their maximum consumption level.
Mathematically, this implies that we have
∂U(x, ω)
∂x
≥ 0. (3)
For notational convenience we define
V (x, ω)
∂U(x, ω)
∂x
, (4)
as the marginal benefit [3], [4].
2) Property II: The marginal benefit of customers is a non-
increasing function and we have
∂V (x, ω)
∂x
≤ 0. (5)
In other words, the utility functions are concave and the level
of satisfaction for users can gradually get saturated. While the
class of utility functions that fulfill (3) and (5) is very large,
it is convenient to have a linear marginal benefit [3], [4].
3) Property III: We have to be able to rank the customers
based on their utilities. In our formulation, we assume, for a
fixed consumption level x, a larger ω implies a larger U(x, ω),
which can be expressed as
∂U(x, ω)
∂ω
> 0. (6)
4) Property IV: We assume the general expectation that no
power consumption brings no benefit, so we have
U(0,ω)=0, ∀ω>0. (7)
Various choices of utility functions are widely used in the
communications and networking literature [19]. However, re-
cent reports indicate that the behavior of power users can also
be accurately modeled by certain utility functions [3]. In this
paper, we consider quadratic utility functions corresponding
to linear decreasing marginal benefit [5]:
U(x, ω)=
⎧
⎨
⎩
ωx −
α
2
x
2
if 0 ≤ x ≤
ω
α
,
ω
α
if x ≥
ω
α
,
(8)
where α is a pre-determined parameter. Sample utility func-
tions from this class are shown in Fig. 1.
A subscriber that consumes x kW electricity during a
designated number of hours at a rate of P dollars per kWh is
charged Px dollars per hour. Hence, the welfare of each user
can simply be represented as
W (x, ω)=U(x, ω) − Px, (9)
where W (x, ω) is the user’s welfare function, U (x, ω) is
the utility function of the user, Px is the cost imposed by
the energy provider to the user, and x is the user’s power
consumption. For each announced price value P , each user
tries to adjust its power consumption x to maximize its own
welfare, and this can be achieved by setting the derivative of
(9) equal zero which means that at the optimal consumption
level, the marginal benefit of the user would be equal to the
announced price. For example, different power consumption
responses of a user with a decreasing linear marginal benefit
to two different announced prices are depicted in Fig. 2.
B. Energy Cost Model
We consider a cost function C
k
(L
k
) indicating the cost of
providing L
k
units of energy offered by the energy provider
in each time slot k ∈K. We make the following assumptions:
Assumption 1: The cost functions are increasing in the
offered energy capacity. That is, for each k ∈K,wehave
C
k
(
ˆ
L
k
) ≤ C
k
(
˜
L
k
), ∀
ˆ
L
k
≤
˜
L
k
. (10)
Assumption 2: The cost functions are strictly convex.For
each k ∈K,any0 ≤ θ ≤ 1, and
ˆ
L
k
,
˜
L
k
≥ 0, we have [20]
C
k
(θ
ˆ
L
k
+(1− θ)
˜
L
k
) ≤ θC
k
(
ˆ
L
k
)+(1− θ)C
k
(
˜
L
k
). (11)
Piece-wise linear functions and quadratic functions are two
example cost functions that satisfy Assumption 1 and Assump-
tion 2. In this paper, we consider quadratic cost functions [10]:
C
k
(L
k
)=a
k
L
2
k
+ b
k
L
k
+ c
k
, (12)
0 0.5 1 1.5 2
0
0.5
1
1.5
Amount of Power Consumption (kW)
Utility / Marginal Benefit
Utility Function (ω = 1)
Utility Function (ω = 0.5)
Utility Function (ω = 0.3)
Marginal Benefit Function (ω = 0.3)
V(x,ω)
U(x,ω)
Fig. 1. Sample utility functions for power subscribers (α =0.3).
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.05
0.1
0.15
0.2
0.25
Amount of Power Consumption (kW)
V(x,ω)
Marginal Benefit Function (ω = 0.25)
Fig. 2. Different power consumption reactions of a subscriber to two different
announced prices (P
1
=0.2, P
2
=0.1 and α =0.25).
where a
k
> 0 and b
k
,c
k
≥ 0 are pre-determined parameters.
III. R
EAL-TIME PRICING FORMULATION
In this section, we formulate the interactions between the
users and the energy provider as an optimization problem and
analyze the existence and uniqueness of the solution. In our
model, the energy provider announces the price of electricity
in real-time based on the total load demand.
A. Optimization Problem Formulation
From a social fairness point of view, it is desirable t o utilize
the available capacity provided by the energy provider in such
a way that the sum of the utility functions of all subscribers
is maximized and the cost imposed to the energy provider is
minimized. However, each subscriber will choose its consump-
tion level to maximize its own welfare function introduced in
(9). These individually optimal consumption levels may not be
socially optimal for a general price announced by the energy
provider. To align these individual optimal consumption levels
with the social optimal case, we need to adopt the sum of
all utility functions minus the cost imposed to the energy
provider as the objective function while the consumption levels
of all users are coupled via the limited available generation
capacity. Having a centralized control over all subscribers,
and also being provided with complete information about the
subscribers’ needs, an efficient energy consumption schedule
can be characterized as the solution of the following problem:
maximize
x
k
i
∈I
k
i
,L
min
k
≤L
k
≤L
max
k
,
i∈N ,k∈K
k∈K
i∈N
U(x
k
i
,ω
k
i
) − C
k
(L
k
)
subject to
i∈N
x
k
i
≤ L
k
, ∀k ∈K,
(13)
where U (x
k
i
,ω
k
i
) is defined in (8), C
k
(L
k
) is defined in (12),
and ω
k
i
is the ω parameter of user i in time slot k.
The problem formulated in (13) is a concave maximiza-
tion problem and can be solved using convex programming
techniques such as the interior point method (IPM) [20] in a
central fashion. However, the problem arising in solving (13)
in a central manner is that we need to know the exact utility
function of users. Since it is assumed that the utility parameter
ω
k
i
for each user i ∈N is private, the energy provider may
not have sufficient information to s olve problem (13).
B. Dual Decomposition Approach
We notice that (13) can be solved independently for each
time slot k ∈K. In other words, we have the following
optimization problem for each fixed time slot k ∈K:
maximize
x
k
i
∈I
k
i
,i∈N ,L
min
k
≤L
k
≤L
max
k
i∈N
U(x
k
i
,ω
k
i
) − C
k
(L
k
)
subject to
i∈N
x
k
i
≤ L
k
.
(14)
Problem (14) is again convex and can be solved easily in a
centralized manner. In practice, this problem has to be solved
in a distributed fashion. Although the objective function in
(14) is further separable in x
k
i
and L
k
, the variables x
k
i
and L
k
are coupled by the imposed constraint that the total consumed
power cannot exceed the available capacity in (14).
For primal problem (14), the Lagrangian is defined as [20]:
L(x,L
k
,λ
k
)=
i∈N
U(x
k
i
,ω
k
i
) − C
k
(L
k
)
−λ
k
(
i∈N
x
k
i
− L
k
),
=
i∈N
(U(x
k
i
,ω
k
i
) − λ
k
x
k
i
)
+λ
k
L
k
− C
k
(L
k
),
(15)
where λ
k
is the Lagrange multiplier and x =(x
k
i
,i∈N) for
afixedk ∈K. Due to the separability of the first term in the
Lagrangian, we can write the objective function of the dual
optimization problem as [20]:
D(λ
k
)= maximize
x
k
i
∈I
k
i
,i∈N ,L
min
k
≤L
k
≤L
max
k
L(x,L
k
,λ
k
)
=
i∈N
B
k
i
(λ
k
)+S
k
(λ
k
),
(16)
where
B
k
i
(λ
k
)=maximize
x
k
i
∈I
k
i
U(x
k
i
,ω
k
i
) − λ
k
x
k
i
, (17)
and
S
k
(λ
k
)= maximize
L
min
k
≤L
k
≤L
max
k
λ
k
L
k
− C
k
(L
k
). (18)
The dual problem is
minimize
λ
k
>0
D(λ
k
). (19)
The first term in D(λ
k
) in (16) can be decomposed into N
separable subproblems in form of (17), which can be solved
by the users, and another subproblem in form of (18), which
can be solved by the energy provider.
We can show that strong duality holds, and we can solve the
dual problem (19) instead of the primal problem (14). In this
case, we can obtain the solution of t he dual problem λ
k∗
, and
each individual subscriber and also the energy provider can
simply solve their own local optimization problem determined
by (17) and (18) to obtain x
k∗
i
and L
∗
k
, respectively.
The key idea which motivates us to propose a real-time
pricing algorithm can be understood if we compare the local
problem (17) that has to be solved by each individual user
with (9), introducing each user’s welfare. In fact, if the energy
provider would be able to charge the users at a rate P = λ
k∗
,
and each individual user tries to maximize its own welfare
function, it will be guaranteed by strong duality that the total
power consumption will not exceed the provided capacity.
C. Distributed Algorithm
We explained in the previous section that by charging the
users with the solution of the dual problem λ
k∗
, we can
achieve the solution of primal problem (14). Interestingly, it
is possible to solve the dual problem in an iterative manner
using the gradient projection method, and in this case we have
λ
k
t+1
=[λ
k
t
− γ
∂D(λ
k
t
)
∂λ
k
]
+
=[λ
k
t
+ γ
i∈N
x
k∗
i
(λ
k
t
) − L
∗
k
(λ
k
t
)
]
+
,
(20)
where t ∈T, and T is the set of time instances at which
the energy provider updates λ
k
. Here, x
k∗
i
(λ
k
t
) is the local
optimizer of (17), and L
∗
k
(λ
k
t
) is the local optimizer of (18)
for a given λ
k
t
, respectively. Also, λ
k
t
is the value of λ
k
in
instance t ∈T, and γ is the step size. The interaction between
the energy provider and the subscribers is depicted in Fig. 3.
The distributed algorithms of each subscriber and the energy
provider are summarized in Algorithms 1 and 2, respectively.
Consider Algorithm 1. In Line 1, each subscriber starts with
its initial condition, which is assumed to be random. Then, the
loop in Lines 2 to 6 describes the responses of each subscriber
to the newly announced price λ
k
. Within this loop, each
subscriber receives the new value of λ
k
in Line 3 and solves
local problem (17) to get the optimal consumption x
k∗
i
(λ
k
)
corresponding to the new value of λ
k
in Line 4. In Line 5, the
user communicates the new value of x
k∗
i
(λ
k
) to the energy
provider. We note that in each time slot k ∈K, users apply
their new loads only after the algorithm has converged.
In Algorithm 2, the energy provider starts with random
initial conditions in Line 1. The loop in Lines 2 to 11 continues
during the operational cycle of the system. Within this loop,
the energy provider updates λ
k
in each instance t ∈T in Lines
4 and 5. It further calculates the new value of L
k
(λ
k
) which
Fig. 3. Illustration of the operation of the proposed algorithm and the
interactions between the energy provider and subscribers in the system.
Algorithm 1 : Executed by each subscriber i ∈N.
1: Initialization.
2: for each t ∈T
3: Receive the new value of λ
k
from energy provider.
4: Update the consumption value x
k∗
i
(λ
k
) by solving (17).
5: Communicate the updated x
k∗
i
(λ
k
) to energy provider.
6: end for
maximizes its welfare and updates its information about the
total consumption level of the system in Lines 7 to 9.
We note that network utility maximization has already been
applied successfully in computer networking. The problem
formulation in this section is similar to the congestion control
problem in the Internet (e.g., [19]). However, the pricing
algorithm in this paper differs from the rate allocation problem
for the Internet in two aspects: (a) The capacity can be adjusted
by the energy provider and may change periodically while the
capacity constraint in [19] is fixed; (b) We consider the energy
cost imposed to the energy provider and formulate the problem
as utility maximization together with cost minimization.
IV. P
ERFORMANCE EVA L UAT I O N
In this section, we present simulation results and assess
the performance of our proposed distributed algorithm. In our
simulation model, we assume there are N =10subscribers.
The entire time cycle is divided into 24 time slots representing
the 24 hours of the day. The minimum and the maximum
power requirements of all users vary in each time slot, and
the minimum generating capacity to meet the minimum power
requirements is guaranteed. However, we also assume the
maximum generating capacity L
max
k
is equal to the maxi-
mum total power requirements of all the users, so we have
L
max
k
=
i∈N
M
k
i
, for all k ∈K.
We also assume the ω parameter of each user is selected
randomly from the interval [1 , 4] and remains fixed within the
entire cycle. Parameter α of the utility function introduced in
Algorithm 2 : Executed by the energy provider.
1: Initialization.
2: repeat
3: if time t ∈T
4: Compute the new value of λ
k
using (20).
5: Broadcast the new value of λ
k
to all the subscribers.
6: else
7: Update the capacity value L
k
(λ
k
) by solving (18).
8: Receive x
k∗
i
(λ
k
) from all the subscribers i ∈N.
9: Update the total load
i∈N
x
k∗
i
(λ
k
) accordingly.
10: end
11: until end of intended period.
(8) is chosen to be 0.5, and we set the parameters of the cost
function introduced in (12) to a
k
=0.01, b
k
=0, and c
k
=0.
Simulation results for the total consumed power for the pro-
posed algorithm are shown in Fig. 4. As illustrated in Fig. 4,
due to real-time interaction of the subscribers and the energy
provider, the two curves corresponding to the total power
consumption of the users and the desired generating capacity
of the energy provider coincide. The high utilization of the
available resources while keeping the total power consumption
below the desired threshold is one of the advantages of the
proposed algorithm. As expected, the generating capacity and
also the total power consumption are bounded within the
minimum and the maximum total power requirements of all
the users in each time slot.
To have a baseline scheme for comparison with the proposed
real-time pricing strategy, we also consider a fixed pricing
scenario with a hard constraint to keep the total consumption
below the generating capacity without interaction with the
users. In the fixed pricing algorithm, the energy provider
announces a price for each time slot k ∈Kat the beginning
of the time slot which guarantees for any type of users with
different choices of the ω parameter that the total consumption
level will not exceed the generating capacity. Therefore, in
the fixed pricing algorithm, the worst case situation where
the ω parameter of all the users assumes the maximum value
ω
max
=4is being considered. Hence, the price in each time
slot k ∈Kcan be calculated as
P
k
fixed
= ω
max
−
L
k
α
N
. (21)
Simulation results for the aggregate utility of all users for the
two different methods are shown in Fig. 5. We can see that
the aggregate utility is much higher for our proposed real-time
pricing algorithm than for the fixed pricing algorithm.
Last but not least, our proposed distributed real-time pricing
algorithm can also benefit the users. Let us consider 24 time
slots with different power requirements for different users
in each time slot. Simulation results for the time averaged
welfare of each individual subscriber for our proposed real-
time pricing algorithm as well as the fixed pricing algorithm
are shown in Fig. 6. We can see that the average welfare of
each individual subscriber is much higher for our proposed
