1
A Decentralized Trading Algorithm for an Electricity Market
with Generation Uncertainty
Shahab Bahrami? and M. Hadi Amini†
?
Department of Electrical and Computer Engineering, The University of British Columbia, Vancouver, Canada
†
Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA, USA
email: bahramis@ece.ubc.ca, amini@cmu.edu,
Abstract—The uncertainties of the renewable generation units
and the proliferation of price-responsive loads make it a challenge
for independent system operators (ISOs) to manage the energy
trading market in the future power systems. A centralized energy
market is not practical for the ISOs due to the high computational
arXiv:1705.02577v1 [cs.SY] 7 May 2017
burden and violating the privacy of different entities, i.e., load
aggregators and generators. In this paper, we propose a day-
ahead decentralized energy trading algorithm for a grid with
generation uncertainty. To address the privacy issues, the ISO
determines some control signals using the Lagrange relaxation
technique to motivate the entities towards an operating point
that jointly optimize the cost of load aggregators and profit of
the generators, as well as the risk of the generation shortage
of the renewable resources. More, specifically, we deploy the
concept of conditional-value-at-risk (CVaR) to minimize the
risk of renewable generation shortage. The performance of the
proposed algorithm is evaluated on an IEEE 30-bus test system.
Results show that the proposed decentralized algorithm converges
to the solution of the ISO’s centralized problem in 45 iterations.
It also benefits both the load aggregators by reducing their cost
by 18% and the generators by increasing their profit by 17.1%.
Keywords: price-responsive load, generation uncertainty, dis-
tributed algorithm, conditional value-at-risk, trading market.
I. I NTRODUCTION
One goal of the future smart grid is to provide the neces-
sary infrastructure towards integration of renewable generators
(e.g., wind turbine, photovoltaic (PV) panel) as the environ-
mentally friendly alternatives for the fossil fuel-based gener-
ators. Meanwhile, the infrastructure in smart grid facilitates
the participation of demand side in the energy management
programs such as demand response (DR) [1].
The optimal power flow (OPF) analysis is widely used by
independent system operators (ISOs) to optimally dispatch the
available generation portfolio to meet the load demand. There
are challenges to address the OPF analysis in a grid with re-
newable energy generators and price-responsive load demands.
First, addressing the OPF problem by the ISO can violate
the entities’ privacy in a competitive energy market, e.g., by
revealing the demand information of the load aggregators and
cost of the generators. Second, the OPF problem is nonlinear
and nonconvex problem, which is difficult to be solved. Even
by linearizing the problem (e.g., in DC OPF), it can be
computationally difficult in large networks, especially when
the number of decision variables increases by participating
the price-responsive load aggregators in the energy market.
Third, the uncertainty of the renewable generation puts the
generation-load balance at risk, especially when the ISO does
not have access to the accurate historical generation data of
the privately owned renewable generators.
There have been some efforts in the power system opera-
tion literature to tackle the above-mentioned challenges. We
divide the related works into three thread. The first thread
is concerned with privacy issues and focuses on designing
a distributed energy planning programs for a system with
multiple suppliers and multiple users using the evolutionary
game [2], Stackelberg game [3], dual decomposition method
[4], supply bidding mechanism [5], and hierarchical bidding
[6]. These approaches may not be easily implementable in
practice, since they ignore the physical constraints imposed
by the topology and operation of the power network. The
second thread is concerned with including the power flow
constraints in the decentralized system planning procedure
using different techniques such as the primal-dual algorithm
[7], convex relaxation [8]–[10], quadratic programming [11],
alternating direction method of multipliers (ADMM) [12]–
[14], and Lagrange relaxation method [15], [16]. In these
works, however, the uncertainty issues related to integrating
renewable generators remain a challenge. The third thread
is concerned with addressing the uncertainty issues by using
different techniques such as chance constrained optimization
[17]–[19], successive constraint enforcement [20], fuzzy sys-
tems [21], stochastic optimization [22], and risk management
[23], [24]. The proposed models are centralized, and cannot
address the privacy and computational complexity challenges.
In this paper, we focus on extending the works in the third
thread by designing a distributed energy trading algorithm in a
day-ahead electricity market with renewable energy generators
and active participation of the load aggregators in load man-
agement programs. The ISO, load aggregators, and generators
use the smart grid’s communication infrastructure to execute
the proposed distributed algorithm and jointly optimize the
generators’ profit and the load aggregators’ cost. The gen-
erators’ and load aggregators solves their own optimization
problems using some locally available information. Hence, the
privacy of each entity is protected. The main challenge is to
determine proper control signals among the ISO, generators,
and load aggregators that enforce the proposed distributed
algorithm converges to the optimal solution of the centralized
problem for an ISO with complete information.
The main contributions of this paper are as follows:
• Risk Evaluation: Inspired by the works in [23], [24],
we deploy a penalty based on the conditional value-at-
risk (CVaR) into the objective functions of the ISO’s
centralized problem and the generators’ local problems
to address the uncertainties of the renewable generation.
It enables the generators with renewable plants to sell
electricity and gain profit, while limiting the risk of high
generation shortage within a certain confidence level. The
ISO can also reduce the risk of generation-load mismatch.
We discuss how to deal with the difference between the
 2

risk aversion levels of the ISO and generators. profile. In the rest of this section, we will provide the models
• Distributed Algorithm Design: In order to protect the for the load aggregators, generators, and ISO.
privacy of load aggregators and generators, and to address 1) Load Aggregator’s Model: Load aggregator i ∈ N is re-
the computational complexity of the centralized approach, sponsible for managing the load demand of its users. The load
we propose a decentralized algorithm based on the La- demand li,h in bus i in time slot h consists of uncontrollable
grange relaxation method [15], [16] that can be executed (or baseload) demand li,h b c
and controllable demand li,h . The
by the entities in the day-ahead market. A load aggregator controllable demand comprises a set Ai of different types of
to manage the controllable loads of its customers. A controllable loads for residential, commercial, and industrial
generator determine its risk minimizing generation level. users connected to bus i. Controllable load a ∈ Ai has a
Meanwhile, the ISO can meet the power flow constraints scheduling horizon Ha,i ⊆ H that defines the time interval,
in the grid, thereby achieving a triple-win result. in which the load should be scheduled. We further divide the
• Performance Evaluation: Simulation results on an IEEE controllable loads into two types based on their characteristics.
30-bus test system show that the proposed decentralized A controllable load of type 1 have hard deadline. It should
algorithm can converge to the global optimal solution to be operated within the scheduling horizon and turned off in
the centralized problem of the ISO in about 45 iterations. other time slots. Examples include the household’s electric
The proposed algorithm also benefits both the load aggre- vehicle (EV) and production line of an industry. A controllable
gators by reducing their cost by 18% and the generators load of type 2 is more flexible. it can be operated in time slots
by increasing their profit by 17.1% and reducing the peak- out of the scheduling horizon at the cost of high discomfort
to-average ratio (PAR) by 15%. When compared with the for the user, e.g., lighting in households, packing process in
centralized approach, our algorithm has a significantly industries, and air conditioner in commercial buildings. Let
lower computational time. when compared with a central- xa,i,h denote the demand of controllable load a ∈ Ai in time
ized method with AC power flow in different test systems, slot h. Let A1i and A2i denote the set of controllable loads of
our approach has a lower running time at the cost of 3% type 1 and 2 for load aggeragtor i ∈ N , respectively. We have
to 8% error due to the DC power flow approximation.
The rest of this paper is organized as follows. Section II
xa,i,h = 0, a ∈ A1i , h 6∈ Ha,i , (1a)
introduces the models for load aggregators, generators and
ISO. In Section III we propose a decentralized algorithm to xa,i,h ≥ 0, a∈ A2i , h 6∈ Ha,i , (1b)
determine the energy market equilibrium. Section IV provides xmin max
a,i,h ≤ xa,i,h ≤ xa,i,h , a∈ A1i ∪ A2i , h ∈ Ha,i , (1c)
simulation results. Section V concludes the paper. Appendices min max
a ∈ A1i ∪ A2i .
P
Xa,i,h ≤ h∈H xa,i,h ≤ Xa,i , (1d)
A and B can be found in the supplementary document.
Scheduling the controllable loads results in a discomfort
II. S YSTEM M ODEL cost for the users. The discomfort cost for type 1 loads depends
only on the total power consumption deviation from the
Consider a day-ahead energy market with a set N of load
desirable value (e.g., the total charging level of an EV is im-
aggregators and a set M of generators. Each load aggregator
portant). For the scheduled power consumption profile xa,i =
is responsible for managing the load demand on be half
(xa,i,h , h ∈ H) and desirable profile xdes a,i = (xdes , h ∈ H),
of its electricity users. Each generator sells electricity to P a,i,h
the market. The load aggregators and generators use a two-
the discomfort cost Ωa,i (xa,i ) = ωa,i h∈Ha,i (xa,i,h −

2
way communication infrastructure to exchange information xdes
a,i,h ) with nonnegative constant ω a,i is a viable can-
with the ISO. The load aggregators and generators are also didate. The discomfort cost for type 2 loads depends on
connected to each other within the transmission network. The both the amount of power consumption and the time of
ISO monitors the power flow through the transmission lines consuming the power. The discomfort cost Ωa,i (xa,i ) =
des 2 0
P P
over the trading horizon H = {1, . . . , H}, where H is the ω (x
h∈Ha,i a,i,h a,i,h − x a,i,h ) + h6∈Ha,i ωa,i,h xa,i,h with
0
number of time slots with equal length, e.g., one hour. time dependent nonnegative coefficients ωa,i,h and ωa,i,h ,
0
To avoid the potential of confusion, we assume that each bus ωa,i,h  ωa,i,h is a viable candidate.
has either a load aggregator or generator. If a load aggregator Let xi = (xa,i , a ∈ Ai ) denote the profile of power
and a generator are connected to the same bus, we divide that demands over all loads of load aggregator i. Load aggregator
agg
bus into two buses connected to each other through a line i aims to minimize the total Pcost Ci (xi ), which includes the
with zero impedance. If neither load aggregator nor generator discomfort cost Ωi (xi ) = a∈Ai Ωa,i (xa,i ) and the payment
b
is connected to a bus, we add a virtual load aggregator with to the ISO to meet
c
P uncontrollable demand li,h and controllable
zero demand to that bus. It enables us to denote the set of demand li,h = a∈Ai xa,i,h . For i ∈ N , we have
buses by N ∪ M and refer a load aggregator or a generator
Ciagg (xi ) = Ωi (xi ) + h∈H ρi,h li,h
P b c


+ li,h . (2)
by its bus number. We use notation L ⊆ (N ∪ M) × (N ∪ M)
to denote the set of transmission lines. In the energy market, Let Xi denote the feasible space defined by constraints
the ISO provides each entity with the day-ahead market price (1a)−(1d) for load aggregator i. Load aggregator i ∈ N solves
values for trading the electricity. Let ρb,h denote the price in the following optimization problem.
time slot h ∈ H for the entity connected to bus b ∈ N ∪ M.
Let vector ρb = (ρb,h , h ∈ H) denote the day-ahead price minimize Ciagg (xi ) (3a)
xi
 3

β
subject to xi ∈ Xi . (3b) j
pute Cej,h (pren
j,h , αj,h ). For further simplification of the objective
 ren,k +
2) Generator’s Model: Generator j ∈ M sells pconv function (9), we replace the terms ∆j,h (pren j,h , pj,h )−αj,h
j,h of k
its conventional unit’s output power to in times slot h. Let in (7) with the auxiliary variables ηj,h for every sample k ∈ K
pconv = (pconv in time slot h ∈ H. We include θj,h (pj,h − pkj,h ) − αj,h ≤ ηj,h k
j j,h , h ∈ H) denote the conventional unit’s k
generation profile for generator j. The cost of a conventional into the constraint set. We define ηj,h = (ηj,h , k ∈ K). Hence,
unit in time slot h is generally an increasing convex function (7) can be rewritten as
of pconv
j,h [26]. The class of quadratic generation cost functions k
ηj,h
βj
X
conv conv
Cj,h (pj,h ) = aj2 (pconv 2 conv
j,h ) +aj1 pj,h +aj0 is well-known [27].
Cej,h (ηj,h , αj,h ) = αj,h + . (8)
K(1−βj )
A generator can also use renewable units, such as solar and k∈K
wind to benefit from their zero generation cost. Due to the The objective of a generator is to maximize its profit πjgen ,
uncertainty, generator j offers a renewable generation profile which is the revenue from selling electricity in the market
pren
j = (pren
j,h , h ∈ H), such that the risk of deviation from with prices ρj,h , h ∈ H minus the generation cost and
the actual generation profile pbjren is minimized. Specifically, the financial risk CVaR in (8). We define decision vector
to prevent the generators from non-credible high generation ψj = (pconv , pren
j j , (ηj,h , αj,h , h ∈ H)) for generator j. The
offers in the market, the ISO charges generator j with renew- generator’s profit is obtained as follows.
able units by the price θj,h , h ∈ H (cents/kW), when its actual
generation pbj,h ren
is lower than its offer pren
j,h . A generator with
πjgen (ψj ) =
renewable units can use risk measures such as the value-at-risk eβj
X
(pconv ren conv


j,h + pj,h )ρj,h −Cj,h (pj,h )− Cj,h (ηj,h , αj,h ) . (9)
(VaR) and conditional value-at-risk (CVaR) to limit the risk h∈H
of generation shortage. Let ∆j,h (pj,h , pbj,h ) denote a function
The local optimization problem for generator j ∈ M is
that captures the penalty for generation shortage in time slot
h for generator j with renewable units [28]. It is defined as maximize πjgen (ψj ) (10a)
ψj
ren +
∆j,h (pren ren
) = θj,h pren
 
j,h , p j,h − p , h ∈ H, j ∈ M, (4) subject to pmin ≤ pconv max
j,h ≤ pj , h ∈ H, (10b)
bj,h bj,h
j
+
where [·] = max{·, 0}. ∆i,h (prenj,h , p
ren
bj,h ) is a random variable, θj,h (pren
j,h − pren,k
j,h ) − αj,h ≤ k
ηj,h , k ∈ K, h ∈ H, (10c)
ren
since the actual generation pbj,h is a stochastic process. Under
0 ≤ αj,h , ηj,h , h ∈ H. (10d)
a given confidence level βj ∈ (0, 1) and the offered generation
level pren
j,h in time slot h, the VaR for generator j is defined as 3) ISO’s Model: A variety of different objectives can be
the minimum threshold cost αj,h , for which the probability of considered for the ISO in the energy market. In this paper,
generation shortage of generator j being less than αj,h is at we consider the objective of minimizing the social cost with
least βj [28]. That is, the risk of generation shortage of all renewable generators
βj over the planning horizon. The social cost is the generation
pren


VaRj,h j,h = min {αj,h | Pr {∆i,h (·) ≤ αj,h } ≥ βj } . (5) cost of the conventional generators plus the discomfort cost
Due to the non-convexity, it is difficult to minimize the of load aggregators. Regarding the risk minimization, the ISO
VaR. The CVaR is an alternative risk measure, which is can consider the CVaR similar to (4)−(8). Remind that from
convex and can be optimized using sampling techniques. The (4), the penalty for the generation shortage is important for a
CVaR for generator j with renewable units in time slot h is generator. However, from the ISO’s perspective, the value of
defined as the expected value of the generation shortage cost generation shortage of the renewable unit is important. That is,
∆j,h (pren ren from the ISO’s point of view, the cost of generation shortage in
j,h , p
bj,h ) when only the costs that are greater than or
βj time slot h corresponding to generator j is ∆ISO ren
j,h (pj,h , p
ren
)=
equal to VaRj,h (pren
bj,h
j,h ) are considered [28]. That is, ren ren +
n [pj,h − pbj,h ] . Moreover, the ISO may consider a confidence

o
βj βj level βjISO different from the confidence level βj of generator j.

pren ren


CVaRj,h j,h = E ∆j,h (·) ∆j,h (·) ≥ VaRj,h pj,h . (6)

β ISO ISO ISO
Similar to (8), we can formulate Cej,hj (ηj,h , αj,h ) as the
It is possible to estimate the CVaR by adopting sample average
approximate CVaR function that the ISO assigns to generator
approximation (SAA) technique [28]. Samples of the random ISO,k
ren j. Here, ηj,h is the auxiliary variable corresponding to term
variable pbj,h for generator j with renewable units in time slot ren ren + ISO ISO,k
h can be observed from the historical record. Consider the [pj,h − pbj,h ] and ηj,h = (ηj,h , k ∈ K) is the vector of
set K = {1, . . . , K} of K samples of the random variable auxiliary variables for samples k ∈ K of the historical data
ren
pbj,h . Let pren,k denote the k th sample of pbj,h ren
for generator for generator j in time slot h. We have
j,h
j in time slot h. The CVaR in (6) can be approximated by βj ISO X ISO,k
ηj,h
βj eβj ren ISO ISO ISO


CVaRj,h (pren
j,h ) ≈ min Cj,h (pj,h , αj,h ), where
Cej,h ηj,h , αj,h = αj,h + . (11)
αj,h
k∈K
K(1−βjISO )
 ren,k +
βj
X ∆j,h (pren
j,h , pj,h )−αj,h As a netural entity, the ISO uses the DC power flow model
pren


Cej,h j,h , αj,h = αj,h+ . (7) as an acceptable framework to determine the active power flow
K(1−βj )
k∈K through the transmission lines and the generation-load balance
Under the given offer pren
j,h , we can use the historical samples in transmission systems [29], [30]. Let δi,h denote the phase
of the renewable unit’s output power in each time slot to com- angle of the voltage in bus i ∈ N in time slot h ∈ H. We
 4

use vector δh = (δi,h , i ∈ N ) to denote the profile of voltage the entities with sufficient access to the day-ahead energy
phase angles in all buses in time slot h. We define the vector of market to determine their own generation and load demand.
decision variables ψ ISO = (pconv
j , pren ISO ISO
j , ηj,h , αj,h , xi , δh , h ∈ The entities compete with each other to optimize their local
H, j ∈ J , i ∈ N ) for the ISO. The ISO’s objective function is problems (3) and (10). The decision vector of the load ag-

X  X conv βjISO
gregator i is the controllable load profile xi and the decision
f ISO ψ ISO = Cj,h (pj,h ) +ϑc Cej,h ISO ISO



ηj,h , αj,h vector for generator j is ψj . The ISO influences the entities
h∈H j∈M
 by using the nodal prices ρ = (ρb , b ∈ N ∪ M) and the
penalties θ = (θj , j ∈ M) for renewable generators. The
X
+ Ωi,h (xi,h ) , (12)
i∈N goal of the ISO is to determine vectors ρ and θ, such that
the market equilibrium coincides with the optimal solution
where ϑc
is a positive weighting coefficient. Minimizing
of the centralized problem (13). The idea is to formulate
f ISO ψ ISO will enable timely adjustment of control settings
the Lagrange relaxation of the ISO’s problem (13) in order
to jointly reduce the consumers’ discomfort cost and the
to divide problem (13) into several subproblems. Then, we
generators’ generation cost, as well reducing the likelihood of
determine vectors ρ and θ such that the subproblems becomes
high generation shortage. Thus, it can improve the economic
the same as load aggregators and generators problems in (3)
efficiency of the system operation. We formulate the ISO’s
and (10). This approach is viable since Due to problem (13) is
centralized problem as
convex and the constraints are linear. Thus, the strong duality
minimize f ISO ψ ISO


(13a) gap condition (Slater’s condition) is satisfied if a feasible
ψ ISO
solution exists [31, Ch. 5].
subject to constraints (3b) and (10b), i ∈ N , j ∈ M, (13b) Let λaggi,h , i ∈ N , h ∈ H denote the Lagrange multiplier
associated with the equality constraint (13c). Let λgen
X
− li,h = bi,r (δi,h − δr,h ), i ∈ N , h ∈ H, (13c) j,h , h ∈
(i,r)∈L H, j ∈ M denote the Lagrange multiplier associated with
X the equality constraint (13d). Let µr,s,h and µr,s,h , h ∈
pconv
j,h + pren
j,h = bj,r (δj,h − δr,h ), j ∈ M, h ∈ H, (13d)
H, (r, s) ∈ L denote the Lagrange multiplier associated with
(j,r)∈L
  the upper and lower inequalities (13e), respectively. Also, let
br,s (δr,h − δs,h ) ≤ pmax
r,s , (r, s) ∈ L, h ∈ H, (13e) ISO,k
γj,h , j ∈ M, k ∈ K, h ∈ H denote the Lagrange multiplier
pren
j,h − pren,k
j,h − αj,h
ISO
≤ ISO,k
ηj,h , j ∈ M, k ∈ K, h ∈ H, (13f) associated with constraint (13f). We define vector φISO =
0≤ ISO,k ISO,k
αj,h , ηj,h , j ∈ M, h ∈ H. (13g) (λagg gen ISO,k
i,h , λj,h , µr,s,h , µr,s,h , γj,h , i ∈ N , j ∈ M, (r, s) ∈
L, k ∈ K, h ∈ H). We have the following main result.
Constraints (13c) and (13d) are the nodal power balance
Theorem 2 The equilibrium of the energy market coincides
equations. Constraint (13e) is the line flow limit, where pmax r,s with the unique solution to the ISO’s centralized problem in
is the maximum power flow limit for line (r, s) ∈ L and br,s
(13) if and only if for i ∈ N , j ∈ M, h ∈ H the ISO sets
is the admittance of line (r, s) ∈ L. Constraint (13f) is for
ISO,k
auxiliary variable ηj,h corresponding to term [pren
j,h − p
ren +
bj,h ] . ρi,h = −λagg
i,h , (14a)
Problem (13) is an optimization problem with a convex ρj,h = λgen
j,h , (14b)
objective function (remind that CVaR is convex) and linear P ISO,k
k∈K γj,h
 1−β 
constraints. By considering one bus as a slack bus (e.g., bus θj,h =
j
. (14c)
1 with δ1,h = 0, h ∈ H), we have the following theorem.
P ISO,k 1 − β ISO
1 − ϑc + k∈K γj,h j
Theorem 1 If the ISO’s centralized problem in (13) has a The proof can be found in Appendix A. It suggests a decen-
feasible point, then the optimal solution exists and is unique. tralized algorithm to determine the solution to problem (13).
To solve the centralized problem in (13), the ISO needs In a decentralized algorithm, the load aggregators, gen-
the information about the load aggregators’ discomfort cost, erators and ISO interact with each other. Let q denote the
conventional generators’ cost, and renewable units’ historical iteration index. Our algorithm involves the initiation phase and
data samples. However, these information may not be available market trading phase. Algorithm 1 describes the interactions
to the ISO. Also, solving problem (13) can be computationally among the load aggregators, generators, and ISO. Algorithm
difficult in large networks. Instead, we develop a decentralized 1 is based on the projected gradient method. The step size in
algorithm. We show that the ISO can determine the price iteration q of this algorithm is denoted by q .
signals ρb,h for bus b ∈ N ∪ M and penalties θj,h , h ∈ H for Initiation phase: Lines 1 to 4 describe the initiation phase.
generators j ∈ M, such that the competitive equilibrium of Market trading phase: The loop involving Lines 5 to 12
the market coincides with the unique solution of the ISO’s describes the market trading phase, in which the load aggre-
centralized problem in (13). Note that the energy market gators, generators and ISO update their decision variables in
equilibrium corresponds to the optimal solution to problems an iterative fashion. This phase includes the following parts:
(3) and (10) for all entities. 1) Information exchange: In Line 6, each load aggregator
q c,q b
i sends its load profile (li,h = li,h + li,h , h ∈ H) to the ISO
III. E NERGY M ARKET I NTERACTIONS via the communication network. Each generator j sends the
The ISO has no direct control over the generators’ and generation profile pconv,q
j and pren,q
j to the ISO as well.
load aggregators’ behavior. Alternatively, the ISO provides 2) ISO’s update: In Line 7, when the ISO receives the
 5

Algorithm 1 Decentralized Energy Market Trading Algorithm. based algorithm converges for a nonsummable diminishing
1: Set q := 1 and ξ := 10−2 . step size [31, Ch. 5].
2: Each load aggregator i ∈ N randomly initializes its controllable
load profile x1i .
3: Each generator j ∈ M randomly initializes its generation profile IV. P ERFORMANCE E VALUATION
p1j . Each generator j with renewable units initializes vectors α1j
and ηj1 randomly, and provides the ISO with confidence level βj .
In this section, we evaluate the performance of our proposed
decentralized algorithm on an IEEE 30-bus test system with
4: The ISO initializes the voltage angles δb1 , b ∈ N ∪ M and the 6 generators and 21 load aggregators.
vector of Lagrange multipliers φISO,1 . 1) Simulation setup: The data for the test system and the
5: Repeat generators cost functions are given in [32]. Since buses 2, 5,
q
6: Each load aggregator i and generator j, sends li,h , h ∈ H and
and 8 have both generator and load aggregator in the original
pconv,q
j and pren,q
j to the ISO, respectively.
7: ISO updates the voltage angles δb,h q
, h ∈ H, b ∈ N ∪ M, and test system, we add the new virtual buses 31, 32, and 33 for
the vector of Lagrange multipliers φISO,q according to (15a) the load aggregators as shown in Fig. S-1 in Appendix B. The
and (15b). trading horizon is one day with H = 24 one-hour time slots.
8: ISO communicates the updated nodal price ρq+1 and penalties The weight coefficient in the ISO’s objective function (12) is
θ q+1 to the entities. set to ϑc = 2 (without units).
9: Each load aggregator i updates the controllable load profile xqi
by solving its local problem in (3).
To obtain different baseload patterns for the load aggre-
10: Each generator j updates its decision vector ψjq by solving its gators, we use a load pattern for about 5 million consumers
local problem in (10). (which includes residential, commercial, and industrial con-
11: q := q + 1. The step size q is updated. sumers) from Ontario, Canada power grid database [33] from
q q−1
12: Until ||δb,h − δb,h || ≤ ξ, h ∈ H, b ∈ N ∪ M. November 1 to November 21 2016. We scale the load pattern
for each bus, such that the average baseload becomes equal to
60% of the load demand of that bus in [32]. To simulate the
information from the entities, it obtains the updated values controllable load demand of each bus, we randomly generate
q+1
of the voltage angles δb,h for b ∈ N ∪ M in time slot the desirable load profile of 500 to 1000 controllable loads of
h ∈ H and the vector of Lagrange multipliers φISO,q+1 using types 1 and 2 with average demand of 2 to 15 kW at each
ISO
the gradient of the Lagrangian function fLag (ψ ISO , φISO ) (See bus. Limits xmax min
a,i,h and xa,i,h for controllable load a in bus
(S-1) in Appendix A) in iteration q as follows: i are set to ±30% of the desirable demand of that load in
max min
q+1 q
time slot h. Limits Xa,i and Xa,i for controllable load a in
δb,h = δb,h + q ∇δi,h
q
ISO
fLag (ψ ISO,q , φISO,q ), (15a) bus i are set to ±5% of the desirable total demand of that
ISO,q+1 ISO,q q ISO
φ =φ + ∇φISO,qfLag (ψ ISO,q , φISO,q ), (15b) load. The discomfort coefficients ωa,i for controllable load
a of type 1 are randomly chosen from a truncated normal
where ∇ is the gradient operator. The ISO uses the results of distribution, which is lower bounded by zero and has a mean
Theorem 2 to compute the updated values of the nodal prices value of ωiavg = 15 cents/(kWh)2 and a standard deviation
ρq+1 and penalties θ q+1 . In Line 8, the ISO communicates of 5 cents/(kWh)2 . The discomfort coefficients ωa,i,h , h ∈ H
the above control signals to the corresponding load aggregator for controllable load a of type 2 in bus i are randomly chosen
and generator. from a truncated normal distribution, which is lower bounded
3) Load aggregator’s update: When load aggregator i by zero and has a mean value of ωiavg = 15 cents/(kWh)2
receives the control signals ρq+1i,h , h ∈ H from the ISO, in and a standard deviation of 5 cents/(kWh)2 . The discomfort
Line 9, it updates its controllable load profile xqi by solving 0
coefficients ωa,i,h , h ∈ H for controllable load a of type 2 in
its local problem in (3). It is a quadratic program with linear bus i are set to 50 cents/(kWh)2 .
constraints and can be solved efficiently by the load aggregator. We assume that generators in buses 8 and 11 have PV panel
Generator’s update: When generator j receives the control and wind turbine, respectively, in addition to their conventional
signals ρq+1 q+1
j,h and θj,h , h ∈ H from the ISO, in Line 10, it units. To obtain the samples for the output power of the wind
updates its decision vector ψjq by solving its local problem turbine and PV panel, we use the available historical data from
in (10). It is a quadratic program with linear constraints and Ontario, Canada power grid database [33], from November 1
can be solved efficiently by the generator using its local to November 21 2016. For each renewable generator, we scale
information about its conventional and renewable units. down the available historical data, such that the average output
Step size update: We use a nonsummable P∞ diminishing step power over the historical data becomes equal to 4 MW.
size with conditions limq→∞ q = 0, q=1  q
= ∞, and The step size is q = 10+0.2q1
. For the benchmark scenario,
q 2
P∞
q 1
q=1  = ∞. One example is  = a+bq , where a and we consider a system without renewable units and DR pro-
b are positive constant coefficients. In Line 11, the step size gram. We perform simulations using Matlab R2016b in a PC
is updated. In Line 12, the stopping criterion is given. For with processor Intel(R) Core(TM) i7-3770K CPU@3.5 GHz.
stoppin criterion, we have used the convergence of the voltage 2) Load aggregators: Each load aggregator executes Algo-
angles, since they depends on the generation and load level rithm 1 to modify the controllable load demand of its users.
at all buses. Hence, the convergence of the voltage angles Fig. 1 shows the load profile of the load aggregators in buses
implies the convergence of all generators’ and load aggrega- 17, 30, 31 (the load in bus 2), 33 (the load in bus 5) and
tors’ decision variables. The proposed Lagrange relaxation- the benchmark system without renewable generators and DR
 6

program. Peak shaving can be observed in the load profiles. Without DR and renewable generation With DR and renewable generation
40 125
Results for all load aggregators verify that by executing the

Load demand (MW)

decentralized Algorithm 1, the peak load demand is reduced by 20 100

16.5% on average. Since the peak load may occur in different 0 75

times for different load aggregators, we consider the index of 7 am 12 pm 6 pm 12 am 6 am 7 am 12 pm 6 pm 12 am 6 am
15 15
load shift percentage, which is the ratio of the total shifted
load demand to the aggregate demand over 24 hours. In Fig. 10 10
2, the shift load percentage index for different values of the
5 5
average scaling coefficient ωiavg at all load buses are provided. 7 am 12 pm 6 pm 12 am 6 am 7 am 12 pm 6 pm 12 am 6 am

When ωiavg increases in a bus, the load aggregator modifies a Time (hour)

lower amount of controllable loads due to a higher discomfort Figure 1. Load demand profiles over 24 hours in buses 31 (up-left), 32
(up-right), 17 (down-left), and 30 (down-right) with and without DR and
cost of its users, and thereby the load shift percentage index renewable generators.
decreases. Load shifting is performed with the goal of reducing

Load shift percentage (%)

the total cost in (2). Fig. 3 shows that the total cost in (2) 20
of the load aggregators is lower (by about 18%) when the
scaling coefficients ωiavg are lower, since the load aggregators
10
can benefit from the price fluctuations and shift more amount
of load demand to the hours with lower price values.
0
3) Renewable generators: Generator 11 has a PV panel and 3 4 7 10 12 14 15 16 17 18 19 20 21 23 24 26 29 30 31 32 33

generator 13 has a wind turbine in addition to their conven-

tional units. Recall that these generators sell the presumed Figure 2. The values of shift load percentage for different load aggregators
avg
with ωi = 500, 50, 30, and 15 cents/(kWh)2 .
output power of the renewable units in the day-ahead market,
and pay a penalty for generation shortage. In Algorithm 1, the
3000
generators use the historical data of the renewable resources
Total cost ($)

2500
and respond to the control signals (14b) and (14c) to determine
2000
the amount of power that they plan to sell in the market. Fig.
1500
4 shows the historical data and the presumed output power of
1000
the PV panel in bus 11 and the wind turbine in bus 13 for the 3 4 7 10 12 14 15 16 17 18 19 20 21 23 24 26 29 30 31 32 33
confidence levels β11 = β13 = 0.9. The obtained presumed
output power will have the lowest risk in the day-ahead market Figure 3. The load aggregator’s cost with low and high discomfort cost’s
for these generators. Two parameters affect the presumed scaling coefficients.
output power of the renewable generator: the value of weight 10 7.5
Renewable generation (MW)

coefficient ϑc in the ISO’s objective function (12) and the 5

5
values of confidence levels βj (set by generator j) and βjISO 2.5
(set by the ISO). The penalty θj,h in (14c) for the generation 0 0
7 am 12 pm 6 pm 12 am 6 am
0 96 192 288 384 478
shortage increases when ϑc increases. That is, when the risk-
10 7.5
averse ISO puts a higher weight on the risk of generation
6
shortage, it assigns a larger penalty. Fig. 5 (a) (for bus 11) and 6
4.5
Fig. 5 (b) (for bus 13) show that a larger coefficient ϑc enforces 2 3
the generators to offer a lower renewable generation in the 0 96 192 288 384 478 7 am 12 pm 6 pm 12 am 6 am
Time (hour)
market during most of the times. In a similar manner, equation
Figure 4. (left figures) The PV panel and wind turbine historical data samples.
(14c) implies that when βj > βjISO , i.e., generator j is more (right figures) The presumed output power of the PV panel and wind turbine
risk-averse than the ISO, then the ISO reduces penalty θj,h in buses 11 and 13.
1−βj
by factor 1−β ISO < 1. If βj < βjISO , i.e., generator j is more
j The reduction in the PAR can reduce the peak load demand,
risk-taker than the ISO, then the ISO increases penalty θj,h
1−βj and thus the generation cost of the generators. However, the
by 1−β ISO > 1 to limit the likelihood of generation shortage.
j revenue of the generators will decrease due to a lower price
4) Conventional generators: The generators with conven- in the market. Fig. 8 shows the profit (revenue minus cost)
tional units can also benefit from the DR program by reducing for the generators with and without DR program. The profit
the PAR in the aggregate demand. For example, Fig. 6 shows of the generators increase by 17.1% on average with DR. To
that using the DR program makes the generation profile of the complete the discussion, we consider different discomfort cost
conventional units in bus 11 and 13 smoother. In other words, coefficients ωiavg , i ∈ N for the load aggregators and report the
as the demand profiles in different buses change toward a pro- average profit of generators in Fig. 9. It is interesting that the
file with a lower peak demand, the required power generation scenario with the most flexible load demands (i.e., ωiavg = 0)
follows the same trend. In order to quantify the impact of the does not lead to the maximum profit for the generators, though
DR program on the generation profile, we provide the value of the generation cost of the generators are lower in this scenario.
the PAR with and without DR program in Fig. 7. Results show The reason is that a higher load flexibility will flatten the load
the reduction in PAR for the generators by 15% on average. profiles, and thus reduces the price values and the revenue of
 7

c c c c c c
103
PV generation (MW)
=0.2 =0.5 =1 =2 =5 =20 75
10 With DR and renewable generation
Without DR and renewable generation

Profit ($)
7.5 50
5
25
2.5
0 0
7 am 12 pm 6 pm 12 am 6 am 1 2 5 8 11 13
Generator bus number
c c c c c c
=0.2 =0.5 =1 =2 =5 =20 Figure 8. The profit of the generators with and without DR and renewable
8
generation (MW)

generation.
Wind turbine

7
6 103
73

Average profit ($)

5
63
4
3 53
7 am 12 pm 6 pm 12 am 6 am
Time (hour) 43

Figure 5. The presumed output power of (a) the PV panel in bus 11, and 33
0 50 100 150 200 250 300 350 400 450 500
(b) the wind turbine in bus 13 for different values of coefficient ϑc .

70 Figure 9. The average profit of the generators in terms of the average weight
Without DR and renewable generation avg
60 With DR and renewable generation coefficients ωi , i ∈ N .
Generation (MW)

50 1.5
1,5 2,5 7,10 5,10 28,15 6,15 20,20 10,20
40
1
30
7 am 12 pm 6 pm 12 am 6 am 0.5
70 0
60 -0.5
50 -1
40
-1.5
30 0 15 30 45 60 75 90
7 am 12 pm 6 pm 12 am 6 am
Time (hour)
Figure 10. The convergence of phase difference over transmission lines.
Figure 6. The generation profile of the generators (a) 11 and (b) 13.
1250
1.6
1000 Centeralized algorithm
CPU time (s)

Without DR and renewable generation

With DR and renewable generation Decentralized algorithm
750
PAR

1.4
500
1.2 250
0
IEE IEE IEE IEE Po Po
1 E1 E3 E3 lish lish
1 2 5 8 11 13 4-b 0-b E1 23 30
us us 18 00 83 12
Generator bus number -bu -bu wp w
s s p
Figure 7. The PAR in the generation of the generators with and without DR
and renewable generation. Figure 11. The CPU time of the centralized and decentralized algorithms.

the generators. Here, the reduction in the revenue is larger than We use MOSEK solver to solve the ISO’s centralized
the reduction in the costs. The maximum profit is achieved for problem (13). The solution is the same as the decentralized
ωiavg = 28.2 cents/(kWh)2 . approach, but the running time is 35 seconds. To further
elaborate the comparison, we provide the average running
time of Algorithm 1 and the centralized approach for six test
5) Algorithm convergence: We study the required number
systems [32] in Fig. 11, we also use the approach in [24] that
of iterations for convergence, which can be interpreted as
applies semidefinite programming (SDP) to obtain the global
an indicator of the number of message exchanges among
optimal point in a grid using AC OPF analysis and the CVaR
the load aggregators, generators and the ISO. The value of
as the risk measure. In addition to comparing the algorithm
voltage angles depends on all generators’ and load aggrega-
running time, the global optimality of the solution to the AC
tors’ decision variables. Thus, the convergence of the voltage
OPF enables us to quantify the approximation in using the
angles is a viable indicator of the convergence of the decision
DC OPF in our decentralized algorithm. The main difference
variables in all buses of the system. Since the values of
is that the AC OPF includes the network losses, while the DC
the voltage angles of all buses can be added by a constant,
OPF doesn not consider the network losses. Furthermore, the
we illustrate the convergence of the phase angle difference
SDP approach in [24] returns the global optimal solution to
between the voltages of the buses at the end nodes of the lines.
the AC OPF. The calculation results show that the values of
As an example, we provide the convergence of δ1,5 − δ2,5 ,
the ISO’s objective using Algorithm 1 with DC OPF are lower
δ7,10 − δ5,10 , δ28,15 − δ6,15 , and δ20,20 − δ10,20 in Figures 10.
by 3% (in Polish 2383wp) to 8% (in IEEE 300-bus system)
We can observe that 45 iterations are enough for convergence.
than the centralized method with AC OPF and SDP approach
The average running time of the algorithm for different initial
due to the inclusion of losses and optimality of SDP method.
conditions is 19 seconds for 200 random initial conditions.
Whereas, the computation time is much lower in Algorithm 1.
 8

Table I [7] V. R. Disfani, L. Fan, and Z. Miao, “Distributed dc optimal power flow
T HE OPTIMAL VALUE AND AVERAGE CPU T IME FOR THE D ETERMINISTIC for radial networks through partial primal dual algorithm,” in 2015 IEEE
M ULTI - STAGE A LGORITHM AND O UR P ROPOSED A LGORITHM . Power & Energy Society General Meeting. IEEE, 2015, pp. 1–5.
[8] W. Shi, N. Li, X. Xie, C. Chu, and R. Gadh, “Optimal residential demand
Centralized algorithm response in distribution networks,” IEEE Journal on Selected Areas in
Algorithm 1
with AC OPF Comm., vol. 32, no. 7, pp. 1441–1450, Jun. 2014.
Test system f ISO ($) CPU time (s) f ISO ($) CPU time (s) [9] N. Li, L. Gan, L. Chen, and S. Low, “An optimization-based demand
response in radial distribution networks,” in Proc. of IEEE Globecom,
IEEE 14-bus 207,780.8 10 217,991.1 24 Anaheim, CA, Anaheim, CA 2012.
[10] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power
IEEE 30-bus 330,760.2 16 347,980.8 63 flow for smart microgrids,” IEEE Trans. on Smart Grid, vol. 4, no. 3,
pp. 1464–1475, Sept. 2013.
IEEE 118-bus 3,290,813.4 25 3,490,815.8 153
[11] A. G. Bakirtzis and P. N. Biskas, “A decentralized solution to the DC-
IEEE 300-bus 18,152,944.3 42 19,845,152.8 218 OPF of interconnected power systems,” IEEE Trans. on Power Systems,
vol. 18, no. 3, pp. 1007–1013, Aug. 2003.
Polish 2383wp 43,263,027.9 136 44,164,101.3 1,312 [12] T. Erseghe, “Distributed optimal power flow using ADMM,” IEEE
Trans. on Power Systems, vol. 29, no. 5, pp. 2370–2380, Sept. 2014.
Polish 3012wp 64,121,518.3 182 71,426,241.2 1,636 [13] Q. Peng and S. H. Low, “Distributed algorithm for optimal power flow
on a radial network,” in Proc. of IEEE Conf. on Decision and Control,
Dec. 2014, pp. 167–172.
[14] S. Magnússon, P. C. Weeraddana, and C. Fischione, “A distributed
V. C ONCLUSION approach for the optimal power-flow problem based on ADMM and
sequential convex approximations,” IEEE Trans. on Control of Network
In this paper, we proposed a decentralized algorithm for Systems, vol. 2, no. 3, pp. 238–253, Sept. 2015.
energy trading among load aggregators and generators in a [15] J. M. Arroyo and F. D. Galiana, “Energy and reserve pricing in security
power grid. In our model, the ISO sends control signals to the and network-constrained electricity markets,” IEEE Trans. on Power
Systems, vol. 20, no. 2, pp. 634–643, May 2005.
entities to motivate them towards optimizing their objectives [16] G. Hug-Glanzmann and G. Andersson, “Decentralized optimal power
independently, while meeting the physical constraints of the flow control for overlapping areas in power systems,” IEEE Trans. on
power network. We introduced the concept of CVaR to limit Power Systems, vol. 24, no. 1, pp. 327–336, Feb. 2009.
the likelihood of renewable generation shortage in the day- [17] H. Zhang and P. Li, “Chance constrained programming for optimal
power flow under uncertainty,” IEEE Transactions on Power Systems,
ahead market. To evaluate the performance of the proposed vol. 26, no. 4, pp. 2417–2424, 2011.
decentralized algorithm, we used an IEEE 30-bus test system [18] D. Bienstock, M. Chertkov, and S. Harnett, “Chance-constrained optimal
connected to some renewable generators. Results confirmed power flow: Risk-aware network control under uncertainty,” SIAM
Review, vol. 56, no. 3, pp. 461–495, 2014.
that the algorithm converges in 45 iterations. We also evaluated [19] Y. Cao, Y. Tan, C. Li, and C. Rehtanz, “Chance-constrained
the price responsive load profiles and generation values for optimization-based unbalanced optimal power flow for radial distribution
various discomfort costs of the load aggregators, and showed networks,” IEEE transactions on power delivery, vol. 28, no. 3, pp.
1855–1864, 2013.
that the proposed decentralized algorithm can benefit the load [20] R. A. Jabr, “Adjustable robust opf with renewable energy sources,” IEEE
aggregators by reducing their cost by 18%, and the generators Transactions on Power Systems, vol. 28, no. 4, pp. 4742–4751, 2013.
by reducing the PAR by 15% and increasing their profit by [21] V. Miranda and J. Saraiva, “Fuzzy modelling of power system optimal
load flow,” in Power Industry Computer Application Conference, 1991.
17.1%. Our algorithm benefit the ISO by maintaining the Conference Proceedings. IEEE, 1991, pp. 386–392.
privacy issues and a lower computational time compared to [22] A. Nasri, S. J. Kazempour, A. J. Conejo, and M. Ghandhari, “Network-
the centralized approach. When compared with a centralized constrained ac unit commitment under uncertainty: A benders’ decom-
position approach,” IEEE transactions on power systems, vol. 31, no. 1,
method with AC power flow equations, our approach has a pp. 412–422, 2016.
lower running time at the cost of 3% to 8% approximation [23] Y. Zhang and G. B. Giannakis, “Robust optimal power flow with wind
error due to using the DC power flow equations. integration using conditional value-at-risk,” in Smart Grid Communi-
cations (SmartGridComm), 2013 IEEE International Conference on.
IEEE, 2013, pp. 654–659.
R EFERENCES [24] P. Samadi, S. Bahrami, V. Wong, and R. Schober, “Power dispatch
and load control with generation uncertainty,” in Proc. of IEEE Global
[1] S. Kar, G. Hug, J. Mohammadi, and J. M. Moura, “Distributed state Conf. on Signal and Information Processing (GlobalSIP), Dec. 2015,
estimation and energy management in smart grids: A consensus innova- pp. 1126–1130.
tions approach,” IEEE Journal of Selected Topics in Signal Processing, [25] L. Song, Y. Xiao, and M. van der Schaar, “Demand side management
vol. 8, no. 6, pp. 1022–1038, 2014. in smart grids using a repeated game framework,” IEEE Selected Areas
[2] B. Chai, J. Chen, Z. Yang, and Y. Zhang, “Demand response manage- in Comm., vol. 32, no. 7, pp. 1412–1424, July 2014.
ment with multiple utility companies: A two-level game approach,” IEEE [26] T. Li and M. Shahidehpour, “Price-based unit commitment: a case of
Trans. on Smart Grid, vol. 5, no. 2, pp. 722–731, Mar. 2014. lagrangian relaxation versus mixed integer programming,” IEEE Trans.
[3] S. Maharjan, Q. Zhu, Y. Zhang, S. Gjessing, and T. Basar, “Dependable on Power Systems, vol. 20, no. 4, pp. 2015–2025, 2005.
demand response management in the smart grid: A stackelberg game [27] J. Stewart, Calculus, 4th ed. CA: Brooks/Cole Pub Co., 1999.
approach,” IEEE Trans. on Smart Grid, vol. 4, no. 1, pp. 120–132, Mar. [28] R. T. Rockafellar and S. Uryasev, “Optimization of conditional value-
2013. at-risk,” J. of Risk, vol. 2, pp. 21–42, Apr. 2000.
[4] R. Deng, Z. Yang, F. Hou, M.-Y. Chow, and J. Chen, “Distributed real- [29] T. J. Overbye, X. Cheng, and Y. Sun, “A comparison of the AC and DC
time demand response in multiseller-multibuyer smart distribution grid,” power flow models for LMP calculations,” in Proceedings of the 37th
IEEE Trans. on Power Systems, vol. PP, no. 99, pp. 1–11, Oct. 2014. Annual Hawaii International Conference on System Sciences. IEEE,
[5] F. Kamyab et al., “Demand response program in smart grid using supply 2004, pp. 9–pp.
function bidding mechanism,” IEEE Trans. on Smart Grid, vol. 7, no. 3, [30] B. Stott, J. Jardim, and O. Alsaç, “DC power flow revisited,” IEEE
pp. 1277 – 1284, 2016. Trans. on Power Systems, vol. 24, no. 3, pp. 1290–1300, 2009.
[6] M. Parvania, M. Fotuhi-Firuzabad, and M. Shahidehpour, “ISO’s optimal [31] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge
strategies for scheduling the hourly demand response in day-ahead University Press, 2004.
markets,” IEEE Trans. on Power Systems, vol. 29, no. 6, pp. 2636–2645, [32] University of Washington, power systems test case archive. [Online].
2014. Available: http://www.ee.washington.edu/research/pstca.
 9

[33] Independent Electricty System Operator (IESO). [Online]. Available: http://www.ieso.ca