© 2015 IEEE.

Personal use of this material is permitted. Permission from IEEE must be

obtained for all other users, including reprinting/ republishing this
material for advertising or promotional purposes, creating new
collective works for resale or redistribution to servers or lists, or reuse
of any copyrighted components of this work in other works.

Digital Object Identifier: 10.1109/TSG.2015.2409053

 1

Consensus + Innovations Approach for Distributed

Multi-Agent Coordination in a Microgrid
Gabriela Hug, IEEE Member, Soummya Kar, IEEE Member, Chenye Wu, IEEE Member

Abstract—Distributed energy resources and demand side man- purpose of being able to parallelize computation and therefore
agement are expected to become more prevalent in the future improve computation speed, multiple situations exist in which
electric power system. Coordinating the increased number of grid physically distributed computations make sense, including (1)
participants in an efficient and reliable way is going to be a major
challenge. A potential solution is the employment of a distributed when the participating entities do not want to share all of their
energy management approach which uses intelligence distributed operational information with any other entity and (2) when it is
over the grid to balance supply and demand. In this paper, we of importance to ensure that a failure of a single computational
specifically consider the situation in which distributed resources entity, i.e., the central coordinator, will not lead to an inability
and loads form microgrids within the bulk power system in which to control the system.
load is supplied by local generation. A distributed energy man-
agement approach based on the consensus + innovations method For distributed approaches in microgrids, research has most-
is presented and used to coordinate local generation, flexible ly focused on setting up the multi-agent structure, ensuring
load and storage devices within the microgrid. The approach interoperability to allow for plug-and-play capability and
takes advantage of the fact that in the optimum the marginal defining the communication structure. In this paper, we present
costs given as a function of the power output/consumption need an algorithm to be implemented in such a multi-agent structure
to be equal for all the network entities (agents). Solutions for
single time step as well as multi time step optimization including and by which the participants in the microgrid coordinate their
inter-temporal constraints are presented. control settings in a distributed way. The proposed approach
Index Terms—Economic Dispatch, Consensus + Innovations is based on the consensus + innovations method [8] and does
Algorithm, Distributed Optimization, Multi-Step Optimization not require any central coordinator or master agent. Such
an algorithm forms the basis for realizing the plug-and-play
capability of a microgrid. Agents are assigned to nodes to
I. I NTRODUCTION
which generators, loads and/or storage devices are connected.
The trend in the electric power system is to move towards These agents define incremental cost/demand functions and
more and more distributed generation resources, distributed constraints for the local energy production and consumption.
storage capabilities and participation of the load in the gen- The consensus portion of the algorithm facilitates the agree-
eration/demand balancing process. This leads to a significant ment on an incremental price for the energy provided and the
increase in the number of entities in the system which need innovation portion ensures that total generation matches total
to be coordinated and to electric energy often being generated demand [9].
closer to the loads. Hence, a possible structure of the future Prior distributed approaches to schedule generation and/or
electric power system could consist of a number of self- load are mostly based on Lagrangian and Augmented La-
sufficient cells of various sizes which internally coordinate grangian Relaxation [10]. Applications of these methods to
their generation and load but also exchange or trade some model predictive control in electric power systems include
amount of power with neighboring cells. [11]–[13]. The approach presented in this paper is concep-
Such a self-sufficient cell is referred to as a microgrid [1]– tually very different from these decomposition theory based
[3]. Microgrids have the capability to disconnect from the approaches. It is based on obtaining a distributed iterative
main grid if needed and locally supply their loads. Prominent solution of the system of first order optimality equations
examples for microgrids include university campuses, military (KKT conditions) associated with the constrained optimization
microgrids and islands. Similar to the bulk power system, problem. Specifically, by exploiting the special structure of the
the key questions are how to ensure the balance between optimality equations, we show that the problem of obtaining
generation and demand and how to achieve this in the most optimal generator allocations can be reduced to a distributed
cost-effective way. Hence, the main focus of this paper is to restricted agreement problem – at any given stage the optimal
design an efficient energy management system for a microgrid. generator allocations are uniquely determined by a single pa-
There are two fundamentally different approaches for the rameter which coincides with the marginal price of generation
design of such an energy management system. One is to assign at the non-binding generators, i.e., the generators which do
the responsibility of coordinating generation, demand, storage not reach their capacity limits at the optimal allocation. We
and main grid connection to a central entity, e.g., [4], [5]. An propose a consensus + innovations approach to ensure that the
optimization problem is solved at the central location and sig- generators reach an agreement on this parameter.
nals are sent to the individual components. Another approach The most relevant related work with regards to consensus
is based on multi-agent systems in which the decisions are based methods has been presented in [14]–[17] where a
made in a distributed way, e.g., [6], [7]. While any distributed decentralized economic dispatch approach based on the con-
approach can be implemented at a centralized location for the sensus algorithm has been introduced. The two key differences
between our approach and this work are that we do not assign
The authors are with the Department of Electrical and Computer Engi-
neering, Carnegie Mellon University, Pittsburgh, PA, 15213, USA, e-mails: any leading role (for achieving global coordination) to any
[email protected], [email protected], [email protected] of the generators and we consider a multi-step optimization
 2

dC/dP dC/dP dC/dP

including inter-temporal constraints allowing for optimal in-
tegration of storage devices and consideration of generation
ramp rates. Usage of the consensus algorithm for the purpose
of ancillary service provision is presented in [18]. However,
the focus is on ensuring resilience against potential packet
drops and it also uses a coordinator which determines the PG PG PL PL PS PS
total amount of required power. Since then, many distributed
C(P ) C(P ) C(P )
control schemes follow up: e.g., see [19], [20] for applications PL PL
to secondary frequency control. Compared to our earlier work
[9], where we employed the consensus + innovations approach
to derive a distributed economic dispatch algorithm, we extend PS PS
the approach to optimize over multiple time steps enabling
distributed Model Predictive Control and we include flexible PG PG
loads and storage devices as controllable components. (a) (b) (c)
The remaining part of the paper is structured as follows: Fig. 1. Marginal cost functions dC/dP and cost functions C for (a)
Sect. II introduces the problem formulation for single and generator, (b) load and (c) storage.
multi-step economic dispatch. In Sect. III, the proposed
distributed approach based on the consensus + innovations
algorithm is derived. Sect. IV provides simulation results and with P n , P n ≥ 0 for generators, P n , P n ≤ 0 for loads
Sect. V concludes the paper. and P n ≤ 0, P n ≥ 0 for storage devices. These bounds
ensure that cost functions are continuously increasing in the
II. P ROBLEM F ORMULATION considered range. Inflexible loads and “must take” generation
We consider a microgrid with N entities, including dis- such as from solar or wind generation can be modeled by
patchable and non-dispatchable generators, critical/inflexible setting upper and lower limits equal to each other resulting in
and non-critical/flexible loads and storage devices. A storage the corresponding P n variables becoming constants instead of
could be a battery but also plug-in electric vehicles which optimization variables.
are available only intermittently. In this section, we provide While the sign convention of C n (Pn ) for loads does not
the mathematical problem formulation which we will use follow standard practice, it should be noted that given P n < 0
subsequently in the next section to derive the distributed the function used does follow the standard convention of a
algorithm. First, we focus on a single time step formulation concave demand function if defined for the positive space.
and then extend it to include inter-temporal constraints in a For a derivation, the reader is referred to the appendix.
multi-step formulation. The objective is to determine the settings of the components,
i.e., Pn ’s, that maximize the (concave) social welfare given by
N

A. Single Time Step
SW = − Cn (Pn ) (4)
We assign a quadratic cost function to each component n n=1
given by
1 In addition, the power balance
Cn (Pn ) = an Pn2 + bn Pn + cn (1)
2 N

with an , bn , cn ≥ 0 and Pn > 0 if the power is generated Pn = 0 (5)
or injected into the system and P n < 0 if it is consumed or n=1
drawn from the system. For generators, the function reflects and the upper and lower bounds on the power inject-
the costs of producing the power P n whereas for loads, it is ed/consumed as defined in (3) for n = 1, . . . , N need to be
the negative benefit function, i.e. the (negative) amount the fulfilled where N is the total number of components to be
load is willing to pay for power |P n |. For a storage device, coordinated. The Lagrange function for this problem can be
Pn is positive whenever the storage is discharging, i.e., the written as2
function corresponds to the amount the storage is willing to N
 N 
 
accept for the provision of P n , and it is negative whenever the L = Cn (Pn ) − λ Pn
storage is charging which corresponds to acting like a load. n=1 n=1
The constant parameter c n for load and storage is assumed to N N
  
be zero. The marginal cost functions 1 are given by + μn Pn − P n + μn (−Pn + P n ) (6)
. n=1 n=1
dCn (Pn )/dPn = an Pn + bn = λn . (2)
where λ, μn , μn are the Lagrange Multipliers for the power
Marginal cost functions dC n /dPn and cost functions C n (Pn ) balance constraint (5) and the upper lower bounds on the
for generator, load and storage are depicted in Fig. 18 in- power outputs (3), respectively. The first order optimality
cluding lower and upper bounds P n and P n , respectively. conditions lead to constraints of the following form
Consequently, Pn is limited to
an Pn + bn − λ + μn − μn = 0, n = 1, . . . , N (7)
P n ≤ Pn ≤ P n (3)
2 For the sake of getting positive values for the Lagrange Multiplier λ of
1 For simplicity we denote cost and benefit functions both as cost functions. (5), a minus sign was added to both sides of (5).
 3

and the supply/demand balance condition (5). and for storage devices additionally
Hence, at the optimal solution, the marginal costs k

. E n ≤ En,0 − Pn,l ≤ E n , k = 0, . . . , K − 1
dCn (Pn )/dPn = an Pn + bn = λn (8) (16)
l=0
for all system components n for which P n has not reached
the upper or lower limit (μn = 0, μn = 0) have to be equal where Cn,k is the cost function of entity n at time step k, P n,k
to the same value, namely the system price λ ∗ at the optimal is the power output of component n at time step k, and K is
solution, i.e. λn = λ∗ , and the power balance (5) needs to the number of steps considered in the multi-step optimization.
be fulfilled [21]. Components for which the optimal setting is Constraint (13) corresponds to the power balance at each step,
P n , yield a marginal cost as defined in (8) that is lower than (14) is the limit on instantaneous power injection/consumption,
λ∗ and components for which the optimal setting is P n , yield (15) corresponds to ramping limits with lower and upper
a marginal cost that is greater then λ ∗ . ramping limits ΔP n , ΔP n and (16) incorporates the upper
Consequently, assuming that the primal problem admits a and lower limits E n , E n on stored energy E n,k . The initial
feasible solution, the optimal settings at all the system entities values for power injection/withdrawal are denoted by P n,−1
can be uniquely parameterized in terms of the quantity λ ∗ , in and En,0 , respectively.
that, the optimal setting P n∗ at an entity n is given by Furthermore, we add a terminal constraint
 ∗  K−1
∗ λ − bn 
Pn = Pn , (9) En,K = En,0 − Pn,l = Enf (17)
an
l=0
where Pn [·] denotes the projection operator associated with for the storage which ensures that the energy level E n,K at
entity n, i.e., it projects the argument into the feasible solution the end of the optimization horizon is equal to a fixed value
space [P n , P n ]: Enf , e.g. Enf = 0.5 · E n . Adding such a terminal constraint in
 ∗  2
λ − bn λ∗ − bn essence is equivalent to adjusting the cost function parameter
Pn = arg min Pn − . (10) bn for different time steps t at which the multi-step optimiza-
an P n ≤Pn ≤P̄n an
tion takes place.
Now, given the power balance constraint (5), the goal of Similarly to the single step problem, we can formulate the
a distributed algorithm can be formalized as a restricted Lagrange function and derive the first order optimality condi-
agreement problem, in which the entities seek to reach an tions for this optimization problem. From these conditions, it
agreement on the quantity λ ∗ that satisfies (see also [9]): can be derived that the optimal solution is given by
N N  ∗  N
λ − bn 
Pn∗ = Pn = 0. (11) Pn,k = 0, k = 0, . . . , K − 1, (18)
n=1 n=1
an
n=1
Furthermore, note that each entity n is only aware of its and marginal costs λ n,k , k = 0, . . . , K − 1 being equal to
local marginal cost/demand function parameters and capacity the optimal system prices, namely λ ∗k , k = 0, . . . , K − 1,
constraints, and hence, cannot directly solve (11). Hence, for the components n for which none of the constraints (14) –
the need for collaboration through inter-entity information (17) becomes binding. For components with binding inequality
exchange arises, which motivates our distributed algorithm in constraints for some time steps within k = 0, . . . , K − 1, the
Sect. III to determine λ ∗ satisfying (11) at each entity. marginal prices deviate from λ ∗k for these specific time steps
by the value of the Lagrange Multipliers of the corresponding
B. Multiple Time Steps binding constraints. However, denote λ ∗ = [λ∗0 , . . . , λ∗K−1 ],
When introducing storage into the system, it is indispensable then the power outputs P n = [Pn,0 , . . . , Pn,K−1 ] for all
to optimize over multiple time steps concurrently because not components are given by
 ∗  2
only instantaneous power is limited but also energy. In order λ − bn . λ∗ − bn
to take into account these inter-temporal constraints and to P n = Pn = arg min P n − , (19)
an P n ∈Xn an
make efficient use of the available storage capacity, we extend
the single step problem to a multi-step optimization problem. where Xn = {Constraints (14) – (17)}. Note that constraints
This allows us to also include other inter-temporal constraints (16) and (17) are only applicable if component n is a storage
such as ramping limitations on generation output. Hence, the device. If the component is a load and ramping of the load is
problem formulation is given by unlimited, then constraint (15) can be neglected.
K−1
 N
max (−Cn,k (Pn,k )) (12) III. D ISTRIBUTED S CHEDULING
Pn,k
k=0 n=1
It is assumed that generators, loads and storage devices
N
 are connected to nodes in the microgrid and an agent is
s.t. Pn,k = 0, k = 0, . . . , K − 1 (13) assigned to each node. The proposed algorithm by which the
n=1 agents coordinate is based on the consensus + innovations
P n,k ≤ Pn,k ≤ P n,k , k = 0, . . . , K − 1 (14) approach. In this section, we first give a general introduction
to this approach, describe how it is used in the single step
ΔP n ≤ Pn,k − Pn,k−1 ≤ ΔP n , k = 1, . . . , K − 1 (15) optimization and then extend it to the multi-step case.
 4

A. Distributed Decision-Making: Consensus + Innovations In the consensus + innovations method, each agent j main-
(i)
We briefly review the consensus + innovations method tains a local copy ν j of the variable ν which is iteratively
and its variants, a generic approach for solving distributed updated, with i denoting the iteration index, as follows:
decision-making problems in multi-agent networks, e.g. [8]. 1) Update local copy of common variable according to
The decision-making setups that fall under the purview of con-  (i) 
dˆ(i)
(i+1) (i) (i)
sensus + innovations typically involve collaborative distributed νj = νj −βi νj − νl −αi n (22)
information processing such as estimation, optimization and l∈ωj n∈Ωj
control in agent networks, in which each network agent has where αi and βi are weight parameters, ω j is the
a priori access to only local information, such as knowledge communication neighborhood of agent j as prescribed
of model parameters and sensed data, and inter-agent com- by the given inter-agent communication topology, i.e.,
munication (interaction) is restricted to a pre-assigned sparse the subset of network agents with which agent j can
communication graph. Broadly speaking, in the consensus exchange information directly, and
+ innovations architecture, the autonomous network agents 
dˆ(i)
(i)
n = Pn dn (νj ) , n ∈ Ωj ,
or decision-makers engage in local information processing (23)
and neighborhood communication to achieve or optimize the
global decision-making task of interest. where Pn [·] denotes the projection operator onto the
For definiteness in this paper, we restrict the discussion of interval [dn , dn ];
the consensus + innovations method to the distributed restrict- 2) Update dependent variables according to (23) to obtain
dˆn , n ∈ Ωj ;
ed agreement problem in multi-agent distributed networks. (i+1)

Formally, in an information processing network, of J agents, (i+1)

3) Exchange local value ν j with the communication
the restricted agreement problem consists of having the J neighbors in ω j and increase iteration counter i.
agents agree on a common value ν subject to the equality (i)
constraint (restriction) Typical conditions that ensure convergence, i.e., ν j → ν
as i → ∞ for all j with ν satisfying (20)-(21), are as follows
J
 J 
 (see, e.g. [8]):
g(ν) = hj (ν) = dn (ν) = 0 (20)
• The local functions d n (·)’s are sufficiently regular .
3
j=1 j=1 n∈Ωj
• The inter-agent communication network needs to be
and inequality constraints connected4.
• The weight parameters α i , βi are positive and satisfy the
dn ≤ dn (ν) ≤ dn , n ∈ Ωj , j = 1, · · · , J, (21)
following conditions:
J
where j=1 |Ωj | = N for some positive integer N , i.e., the – As i → ∞, the sequences {αi } and {βi } are
sets {Ωj }j∈J constitute a partition of [1, · · · , N ], d n (·), n = decaying, i.e., α i → 0, βi → 0.
1, · · · , N are certain real-valued functions, and dn , dn ∈ – The excitations are persistent, i.e.,
[−∞, ∞], for n = 1, · · · , N , are constants. Moreover, we are  
interested in a distributed solution of the restricted agreement αi = βi = ∞ (24)
problem in which, (i) to start with, each agent j is only aware i≥0 i≥0

of its set of local functions d n (·) : R → R, n ∈ Ωj and the – The consensus potential asymptotically dominates
corresponding inequality constraints (21), and (ii) inter-agent the innovation potential, i.e., β i /αi → ∞ as i → ∞.
communication for information exchange is restricted to a pre- In contrast to many primal-dual distributed algorithms, in
assigned communication graph. Under broad assumptions on our approach, each entity updates its local estimate of this pa-
the local functions d n (·)’s and the inter-agent communication rameter by embedding in the same iteration a local innovation,
topology (to be made precise later) an iterative algorithm of which plays the role of a gradient based on its local cost and
the consensus + innovations type may be applied to solve the local knowledge of the power balance equation, and, simulta-
above distributed restricted agreement problem. neously a local agreement or consensus term computed from
Before proceeding to the general consensus + innovations information (parameter estimates) received from its communi-
solution, as a side remark, we comment on a specific instance cation neighbors. Mathematically, both our approach and the
of the above distributed restricted agreement problem in which primal-dual distributed algorithms utilize the local innovation
the (aggregate) local functions h j (·)’s are affine functions of terms. However, the primal-dual algorithms requires additional
the form hj (ν) = ν − xj where the xj ’s are real constants set of Lagrangian multipliers to perform the consensus while
and dn = −∞ and dn = ∞ for all n ∈ Ωj and j = 1, · · · , J, our approach directly tracks the consensus. The saving of
i.e., the inequality constraints are relaxed. In other words, the the additional set of Lagrangian multipliers comes at the
J
agents want to agree on the average value (1/J) j=0 xj . cost of the time-varying weight sequences associated with
Historically, this version of the restricted agreement problem the consensus and innovation potentials. They need to be
is referred to as the average consensus problem or simply designed carefully. More precisely, the consensus potential
the agreement problem and has been studied extensively over
the last few years, e.g. see the review papers [22], [23]. 3 Several regularity conditions of the form of Lipschitz continuity, mono-

The problem with generic functions d n (·)’s and inequality tonicity etc. are presented in [8] that ensure convergence.
4 By connectivity, we mean that there exists a path, possibly multi-hop,
constraints is more involved and in the following we describe
between any pair of agents. As such, the communication network may be
the consensus + innovations method [8] for a general solution quite sparse and in fact, much sparser, than the physical grid topology which
methodology. is typically dense.
 5

should asymptotically dominate the innovation potential, in pair of agents, i.e., the inter-agent communication network is
order to achieve convergence. From a technical viewpoint, the connected. Moreover, note that, in the update rule (26) each
intermixing between the two potentials makes the distributed agent needs to be aware of its local model parameters (a n ’s and
algorithm and its convergence analysis fundamentally different bn ’s) only; this is unlike a centralized optimization approach
from classical gradient based iterative approaches. In addition, in which each agent needs to communicate and coordinate
algorithms of the consensus + innovations type are robust with a fusion center, the latter having access to the model
to a wide class of perturbations resulting from intermittent parameters of all the agents. Interestingly enough, even under
inter-agent communication failures, noisy or quantized data such restrictions on agent communication and lack of global
exchange, and other forms of randomness in the communica- model information, the proposed update rule converges to the
tion infrastructure. For a discussion on the robustness of the optimal schedules at each agent (see [9]).
algorithm the reader is referred to [8].
C. Multi-Step
B. Single Step In the multi-step application as described in Sect. II-B, a
In the single step application, only the scheduling for one horizon of multiple time steps in the future is considered
time step is considered. To this end, let j = 1, · · · , J denote concurrently. This results in an overall decrease in the cost
the nodes (agents) in the microgrid and let n ∈ Ω j index the of supply as preventive actions such as charging the storage
components (generators, loads and storage devices) connected device in anticipation of needed energy in the future can be
to node j and N = n∈Ωj |Ωj |. The common variable ν that taken. In this case, the consensus + innovations application
the agents need to agree on corresponds to the marginal cost of becomes multi-dimensional, i.e., λ j and P n , n ∈ Ωj are
supply λ for that time step whereas the constraint g(ν) which vectors including the prices/marginal costs and the power
needs to be fulfilled is the power balance between supply and output/consumption for the time steps within the horizon,
demand. Specifically, according to (11) and the development respectively. The update given in (26) still applies, but now
in Sect. III-A, the local component function dˆn (λ) is the power as a vector update resulting in the following update for each
Pn injected/drawn by component n which can be given as a time step k
function of λ by
(i+1) (i)
 (i) (i)
 (i)
  λj,k = λj,k − βi λj,k − λl,k − αi Pn,k (28)
ˆ λ − bn
dn (λ) = Pn (λ) = Pn . (25) l∈ωj n∈Ωj
an
(i+1) (i+1) (i+1)
According to (22), each agent j carries out the following with k = 0, . . . , K −1. Denote λj = [λj,0 , . . . , λj,K−1 ].
iterative calculations Then, the power output P (i+1)
n can be updated by
(i+1) (i)
 (i) (i)
  (i+1) 
λj = λj − βi λj − λl − αi Pn(i) (26) λj − bn
l∈ωj n∈Ωj
P (i+1)
n = Pn , ∀n ∈ Ωj , j = 1, . . . , J, (29)
an
Hence, the part of the innovation term assigned to agent j
where the projection operation P n is defined in (19). The same
corresponds to the sum of power injections/consumptions of
statements concerning communication topology and connec-
the components connected to node j, i.e., n ∈ Ω j .
(i+1) tivity hold as in the single step application.
Then, the power injected/consumed P n for n ∈ Ωj is
updated according to
 (i+1)  D. Overview Application Flow
(i+1) λ − bn
Pn = Pn The application of the multi-step consensus approach is
an visualized in Fig. 2. At time t = t 1 the goal is to determine
2 (27)
. λ(i+1) − bn the optimal settings Pn,k for the generators, loads and storage
= arg min Pn − , devices for the next k = 0, . . . , K − 1 steps fulfilling all
Pnmin ≤Pn ≤Pnmax an
constraints on ramp rates, maximum generation output, energy
where Pnmin and Pnmax are defined as in Table I. Hence, level, etc. Hence, the optimization horizon corresponds to
the upper and lower capacity constraints are adjusted to take t = t1 , . . . , t1 + K. The consensus algorithm as described
into account generation ramp rate and energy level constraints in Sect. III-C is used to determine these settings with i
for this one single time step. The initial generator output corresponding to the iteration counter. Once the agents have
and storage energy level are denoted by P n,−1 and En,0 , agreed on λj,k , k = 0, . . . , K − 1 and optimal settings
respectively, and T is the length of one time step. Pn,k , k = 0, . . . , K − 1 have been found, the first step P n,1 is
Consequently, the projection into the solution space in the applied, the optimization horizon is moved by one time step
single step application simply corresponds to limiting the and the consensus algorithm is restarted for t 2 .
power output to values within the given bounds.
Note that the inter-agent communication topology (i.e., who TABLE I
I NCORPORATION OF INTER - TEMPORAL CONSTRAINTS INTO UPPER AND
talks to whom) may differ significantly from the physical pow- LOWER BOUND FOR POWER INJECTION / WITHDRAWAL .
er system electrical coupling among the agents and is, in fact,
Pnmin Pnmax
typically much sparser than the physical coupling topology. In
Gen. max(P n , Pn,−1 + ΔP n ) min(P n , Pn,−1 + ΔP n )
particular, we only assume that there exists a path comprising,
Load Pn Pn
perhaps, of multiple communication hops between any two Stor. max(P n , T1 (En,0 − E n )) min(P n , T1 (En,0 − E n ))
 6

i
the very first time step t = 0, no solution λ t−1 j , P t−1
n for
the previous time step is available to use as starting point in

Consensus
(i) (i)
t = t1 λj , P n the consensus algorithm. Hence, an initialization is required,
e.g. by using a flat start or by using historic solution data.
0
Optimization Horizon In the following time steps t, λ t−1
j , P t−1
n corresponds to the
t1
t solution at the previous step t − 1. An advantage of the multi-
k step application is that good initial settings for k = 0, . . . , K −
0 Pn,k K 1 are available from the solution at the previous time step.
λj,k Only the initial settings for the newly included time step k =
K need to be found. We make the assumption that the settings
i for k = K will be close to the settings at k = K − 1 resulting
in initial settings
Consensus

(i) (i)
λj , P n (0)
t = t2 λj = [λt−1 t−1
j,2...K , λj,K ] (30)

0
Optimization Horizon P (0)
n = [P t−1 t−1
n,2...K , P n,K ] (31)
t
t2 for the consensus algorithm. Then, the iterative process of the
k
0 Pn,k K consensus approach is carried out, i.e., λ j , P n are updated and
λj,k after each iteration the stoping criteria is checked. If changes
Fig. 2. Visualization of multi-step consensus algorithm. in λj and P n are lower than  1 and 2 , respectively, the
consensus iterations are stopped and restarted for a shifted
optimization horizon with initial values given by (30) and (31).
The choice of the initial starting point in the iterative process
For the single step approach, the procedure is very similar
of the consensus algorithm has significant influence on the
to setting the horizon equal to K = 1. In that case, λ j , Pn
convergence of the algorithm. In the considered application for
are scalars and the initial values for the consensus algorithm
the coordination in a microgrid, it can be assumed that most
correspond the values obtained at the previous time step.
of the time the external inputs such as the parameters a n , bn
and any upper and lower limits do not change drastically from
one time step to the next. Hence, a reasonable approach is to IV. S IMULATIONS
use the solution determined at the previous time step t − 1 as In this section, simulation results are provided as a proof of
a starting point for the time step t. concept. Due to limited space, we focus on simulations for the
Figure 3 gives an overview over the entire application flow multi-step application. For single step simulations, the reader
where the solution for time step t is denoted by λ tj , P tn . In is referred to [9].

A. Simulation Setup
Set t := 0
Initialize λt−1
j , P t−1
n The considered microgrid consists of 14 nodes as shown in
Fig. 4. Dispatchable generators are located at buses 1, 2, 3, 6
and 8, a wind generator is placed at bus 9 and a storage device
Set i := 0 is placed at bus 7. The connections indicate communication
(i) lines between components in the microgrid. The test system
λj = [λt−1 t−1
j,2...K , λj,K ]
(i) has been derived from the IEEE 14 bus system using the
Pn = [P t−1 t−1
n,2...K , P n,K ] physical connections as indications for communication lines.
However, physical and communication connections do not
(i+1)
necessarily have to coincide.
Update λj The parameters for generators, loads and storage are given
Single-Step: (26), Multi-Step: (28)
in Tables II - IV. It is assumed that the minimum generation

W
(i+1) 12 13 14 9 8
Update P n
Single-Step: (27), Multi-Step: (29)

11 10
(i+1) (i)
|λj − λj | < 1 ?
6 7
(i+1) (i) no
|P n − P n | < 2 ?
i := i + 1 1 5 4
yes

Solution for t:
λtj = λj
(i+1) t := t + 1
(i+1)
P tn = P n 2 3

Fig. 4. Test System.

Fig. 3. Flow chart consensus update.
 7

TABLE II 2.0
G ENERATOR PARAMETERS ( MU = MONETARY UNITS ) actual
predicted

Load Mult. Factor

Bus an [mu/pu2 ] bn [mu/pu] ΔP n [pu/5min] P n [pu] 1.5
1 0.084 2.0 20 100
2 0.056 3.0 5 105 1.0
3 0.070 4.0 10 100
6 0.060 4.0 5 90
8 0.080 2.5 15 80 0.5

TABLE III 0 50 100 150 200 250

L OAD PARAMETERS ( MU = MONETARY UNITS ) Time Steps (5 min)
Fig. 5. Load Curve.
Bus an [mu/pu2 ] bn [mu/pu] P n [pu] P n [pu]
2 0.080 7.5 -50 -5
3 0.060 6.0 -40 -3 actual
4 0.070 8.0 -80 -6 predicted
30
5 0.064 6.5 -30 -2

Wind [pu]
6 0.040 7.5 -70 -8
9 0.060 8.0 -80 -10 20
10 0.076 7.0 -40 -5
11 0.070 7.5 -25 -2
12 0.080 8.0 -90 -8 10
13 0.070 7.0 -30 -2
14 0.084 8.0 -80 -10 0
50 100 150 200 250
Time Steps (5 min)
TABLE IV Fig. 6. Wind Curve.
S TORAGE PARAMETERS ( MU = MONETARY UNITS )
Bus an [mu/pu2 ] ΔP n [pu] P n [pu/5min] E n [puh]
7 0.02 50 50 20 power output over the entire day is given in Fig. 6. Again, it
is assumed that predictions for the near term future are more
accurate than the predictions for the rest of the optimization
level for all dispatchable generators is 0pu and that the ramp horizon. Hence, again, for the first five time steps k = 0, . . . , 4
up and down limits are equal, i.e., ΔP n = −ΔP n . The in the prediction horizon the prediction errors are reduced
simulations are carried out for an entire day in 5 minute inversely proportionally to k.
intervals and the prediction horizon K is chosen to be equal After trying different combinations, the tuning parameters
to 20 time steps. α and β are set to
It can be assumed that the parameters for the generators
and storage stay the same over a longer time period whereas 0.055 0.2
, αi =
βi = 0.001 (32)
the parameters for the load change to reflect varying needs i0.98 i
over the course of the day. Each load is an aggregation of where i is the iteration counter. Stopping tolerances as defined
multiple loads, including flexible and inflexible loads. Varying in Fig. 3 are set to
needs for such an aggregated load are modeled by adjusting
the parameter b n in the demand function and the minimum 1 = 0.0001, 2 = 0.001. (33)
and maximum demand limits P n and P n , respectively. It is
assumed that the agent of each load predicts the parameters
required to cover its consumption and then schedules the load B. Simulation Results: Normal Operation
according to the outcome of the distributed energy balancing The number of iterations required to converge given the
management. The upper limits P n , i.e., lowest absolute values convergence criteria and simulation parameters defined in the
for the demand, correspond to the portion of the fixed loads. previous section are given in Fig. 7. Depending on the time
As predictive optimization is employed, the agents need to step, the algorithm converges within 50 to 250 iterations with
predict the consumption for the entire prediction horizon K but an average of 140 steps for this particular setup and the chosen
these predictions may be inaccurate. Hence, such uncertainty tuning parameter values.
is simulated by introducing prediction errors. To simulate the For comparison purposes, we have also implemented ADM-
varying demand over the day, we multiply b n , P n , P n with M [13], [24] for the same problem. It converges within 3 to 57
the values given in Fig. 5. To distinguish between prediction iterations with an average of 11 iterations. Note however, that
and actually occurring load, the bold line in Fig. 5 gives the ADMM is by nature a sequential algorithm and each iteration
multiplication factor which actually occurs and the dotted line requires 15 sequential updates in this case. Therefore, though
what has been predicted. Hence, b n , P n , P n are multiplied it may overlook many aspects, our approach is comparable
with the values of the dotted line for the predictions and with ADMM in terms of iterations. To quantitatively illustrate
multiplied with the values of the bold line for the actual the performances, we compare the average execution time
realizations of the load. It is assumed that predictions improve of these two approaches using 64-bit Matlab  R2012a on
the closer the predicted time step to the current time step, i.e., a desktop (Intel  CoreTM i7-4770, 3.46Hz, 32 GB RAM).
for the first five time steps k = 0, . . . , 4 in the prediction We implement both approaches in a sequential way due to
horizon the errors are reduced from what is shown in Fig. 5 the sequential nature of ADMM. That is, the updates at each
inversely proportionally to k. bus are carried out one after another. Our algorithm takes
Similarly, also the output from the non-dispatchable genera- an average of 13s to converge for each execution, while
tor at bus 9 needs to be predicted. The actual and the predicted ADMM takes an average of 88s to converge with the same
 8

20
250
Cons. + Innov.
200 ADMM
15

Energy [puh]
# Iterations

150
10
100

50 5

50 100 150 200 250 0

Time Steps (5 min) 50 100 150 200 250
Time Steps (5 min)
Fig. 7. Number of iteration steps required to converge for Consensus + Fig. 10. Energy level in the storage.
Innovations approach and for ADMM.

Ramping [pu/5min]
stopping criteria. Note that, the convergence of our approach
4
can be sped up by parallel computation and also by an
2
enhanced communication network as will be discussed in Sect. 0
IV-D. Furthermore, it does not require a global update of -2
the Lagrange multiplier associated with the power balance -4
equation as is the case in ADMM. Consequently, the results -6
for our approach are very promising.
High accuracy has been chosen as convergence criterion. 50 100 150 200 250
Time Steps (5 min)
If accuracy is to be improved even further, more iteration Fig. 11. Ramping of generators.
steps will be necessary. On the other hand, if accuracy can
be reduced, also the number of required iterations reduces. It
should be noted that the same settings for α i and βi as defined Figure 8 shows the generation and the storage power output
in (32) are used for each time step t. The values decay over settings over the course of the simulated time range. It can be
the iterations but the initial values and the speed of decay is seen that the proposed algorithm is capable of successfully
the same for every simulated time step. Hence, making these taking into account upper and lower limits on generation
settings adaptive for different levels of loading based e.g. on capacity. Figure 9 provides the consumption by the individual
learning would improve convergence. loads. At the beginning and the end of the simulation, there are
If bounded deviation from the optimal solution can be multiple time steps during which loads reach their lower limits
tolerated, it is also possible to further speed up the convergence indicating that only inflexible/fixed loads are supplied at these
by choosing fixed, non-decaying values for α and β. In this nodes. In Fig. 10, the resulting level of energy in the storage
case, the algorithm in fact converges exponentially fast, at the device is given. The storage is mostly used for arbitrage,
cost of a deviation from the optimal solution. The magnitude i.e., charging in low load situations and discharging in high
of the deviation is a function of the chosen values for α and load situations allowing for a reduced required generation
β [9]. Hence, one can also trade off between this deviation capacity. Furthermore, it provides support in balancing out
error and the convergence speed. some of the short term variability. Figure 11 provides the
ramping of the generators, indicating that constraints on ramp
rates are met.
Generators To provide insight into the convergence behavior, Fig. 12
shows the evolution of λ for the consensus iterations at time
Power Output [pu]

80 Storage
step t = 160. The dimension of λ for each node j is equal to
the prediction horizon. Here, we show the λ’s at all nodes for
40
steps k = 1, 8, 14, 20 in the prediction horizon.

0
C. Simulation Results: Generator Disconnect
50 100 150 200 250 An advantage of the fully distributed approach is that it
Time Steps (5 min) allows for a plug-and-play mechanism. Hence, in the following
Fig. 8. Power output from generation and storage. simulation, the performance of the proposed algorithm is tested
for the case in which the generator at node 3 is disconnected
80
at time step t = 80 and reconnected at time t = 90. Figures
13 and 14 give the number of iterations to converge and the
60
generator and storage outputs. It can be seen that convergence
Load [pu]

40
is achieved within ∼ 450 iterations in the time steps of the
disconnection and reconnection. The iterations for the other
20 time steps stay in the range as in Fig. 7. The reason for
the increased number of iterations is the fact that we assume
0 that the disconnection/reconnection has not been predicted,
50 100 150 200 250
Time Steps (5 min) hence, the initial point for the consensus algorithm is not as
Fig. 9. Consumption by the loads. accurate as with no disconnection. The figures clearly show the
 9

6.80 In addition to the communication network shown in Fig. 4,

6.86 we add three communication links, namely between nodes
λ1 [m.u.]

λ8 [m.u.]
3 and 12, between nodes 1 and 8 and between nodes 6
6.79
6.84 and 7. The additional communication links have been chosen
such as to connect buses which are at opposite ends of the
6.82 6.78 communication network with the result that the diameter of
20 40 60 80 20 40 60 80 the network defined as the maximum number of nodes over
Iteration Iteration
which information needs be communicated to travel from one
6.81 node to another in the network is reduced. The effect is that
6.79 information spreads faster in the network leading to improved
λ14 [m.u.]

λ20 [m.u.]
6.80
convergence.
6.78 6.79 The resulting numbers of iterations for the normal operation
case, i.e., the same simulation setup as in Sect. IV-B but with
6.78 additional communication links, are provided in Fig. 17. It can
6.77
40 60 80 20 20 40 60 80 be seen that the required number of iterations has decreased.
Iteration Iteration
Fig. 12. Evolution of λ’s for all nodes for prediction steps k = {1, 8, 14, 20} V. C ONCLUSION
at time step t = 160.
In this paper, an approach for the coordination of agents in
a microgrid is presented. It is assumed that each generator,
disconnect of the generator and the increase in power injection load and storage device is connected to a specific node which
at the maximum ramp rate as soon as it is reconnected. is shadowed by an agent. The respective agent defines cost
Figures 15 and 16 show the evolution of the λ’s over and demand functions for the producers and consumers at
the iteration of the consensus algorithm at the times of the its node and communicates with the agents of neighboring
disconnect and the reconnect. It is obvious that a greater nodes. The goal of the communication process is to find
correction from the initial point is required which leads to an agreement on an incremental “price” for power provision
a higher number of iterations to fulfill the stopping criterion. while ensuring that overall generation is equal to load. The
proposed approach uses the consensus + innovations method
allowing for a robust and fully distributed coordination of
D. Simulation Results: Additional Communication
the components in the microgrid. Optimal usage of available
The convergence rate, measured in number of iterations until storage and incorporation of ramp rate limitations is achieved
a pre-defined convergence criterion is fulfilled, of any distribut- by applying the method to a multi-step economic dispatch. In
ed algorithm is dependent on the choice of the communication the multi-step case, agents need to agree on prices for each
graph. It is not the focus of this paper to derive the optimal time interval in the prediction horizon. Operational constraints
communication graph, however, in this section, we provide are observed by projecting the obtained temporary solution
simulations which indicate how the convergence rate can be into the feasible solution space. Adding a receding horizon
improved by adding additional communication links. to the multi-step optimization enables a distributed Model
Predictive Control approach and allows for good initial starting
500
points for the consensus algorithm. Future work focuses on
400 including power flow constraints, which leads to a multi
variable updating scheme in a multi time step setting.
# Iterations

300

200 9.0
8.0
100 8.5
λ8 [m.u.]
λ1 [m.u.]

7.5
100 50 150 200 250 8.0
Time Steps (5 min) 7.0
Fig. 13. Iteration steps required to converge for generator disconnects at 7.5
t = 80 and reconnect at t = 90
6.5
100 200 300 400 100 200 300 400
Iteration Iteration
Generators
Power Output [pu]

80 Storage
9.5 10.0
λ14 [m.u.]

λ20 [m.u.]

9.5
40 9.0
9.0
8.5
0 8.5
8.0
8.0
50 100 150 200 250 100 200 300 400 100 200 300 400
Time Steps (5 min) Iteration Iteration
Fig. 14. Power output from generation and storage for generator disconnect Fig. 15. Evolution of λ’s for all nodes for prediction steps k = {1, 8, 14, 20}
at t = 80 and reconnect at t = 90. at time step t = 80.
 10

[13] M. Kraning, E. Chu, J. Lavaei, and S. Boyd, “Dynamic network energy

7.8 8.0 management via proximal message passing”, Foundations and Trends

λ8 [m.u.]
λ1 [m.u.]

in Optimization, vol. 1, no. 2, pp. 1–54, 2013.

7.5 [14] Z. Zhang, X. Ying, and M.-Y. Chow, “Decentralizing the economic
dispatch problem using a two-level incremental cost consensus algorithm
7.6 7.0 in a smart grid environment”, in North American Power Symposium
(NAPS), Aug. 2011.
7.5 [15] Z. Zhang and M.-Y. Chow, “Incremental cost consensus algorithm in a
100 200 300 400 100 200 300 400 smart grid environment”, in IEEE Power and Energy Society General
Iteration Iteration Meeting, July 2011.
9.0 [16] Z. Zhang and M.-Y. Chow, “The leader election criterion for decentral-
8.5 ized economic dispatch using incremental cost consensus algorithm”, in
8.5 IECON 2011 - 37th Annual Conference on IEEE Industrial Electronics
λ14 [m.u.]

λ20 [m.u.]
8.0 Society, Nov. 2011, pp. 2730 –2735.
8.0 [17] Z. Zhang and M.-Y. Chow, “Convergence analysis of the incremen-
7.5 tal cost consensus algorithm under different communication network
7.5 topologies in a smart grid”, IEEE Transactions on Power Systems, vol.
7.0 27, no. 4, pp. 1761 –1768, Nov. 2012.
7.0 [18] A.D. Dominguez-Garcia, C.N. Hadjicostis, and N.H. Vaidya, “Resilient
100 200 300 400 100 200 300 400 networked control of distributed energy resources”, IEEE Journal on
Iteration Iteration Selected Areas in Communications, vol. 30, no. 6, pp. 1137 –1148, July
Fig. 16. Evolution of λ’s for all nodes for prediction steps k = {1, 8, 14, 20} 2012.
at time step t = 90. [19] Changhong Zhao, Ufuk Topcu, and Steven Low, “Swing dynamics
as primal-dual algorithm for optimal load control”, in Proc. of IEEE
SmartGridComm 2012. IEEE, 2012, pp. 570–575.
[20] C. Zhao, U. Topcu, N. Li, and S. Low, “Design and stability of load-
150 side primary frequency control in power systems”, To Appear in IEEE
Trans. on Automatic Control, 2014.
# Iterations

[21] Arthur R. Bergen and Vijay Vittal, Power Systems Analysis, Prentice
100
Hall, 2 edition, Aug. 1999.
[22] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and
50 cooperation in networked multi-agent systems”, Proceedings of the
IEEE, vol. 95, no. 1, pp. 215–233, January 2007.
[23] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione,
“Gossip algorithms for distributed signal processing”, Proceedings of
50 100 150 200 250 the IEEE, vol. 98, no. 11, pp. 1847–1864, Nov. 2010.
Time Steps (5 min)
[24] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and Jonathan
Fig. 17. Iteration steps required to converge with additional communication. Eckstein, “Distributed optimization and statistical learning via the
alternating direction method of multipliers”, Foundations and Trends R
in Machine Learning, vol. 3, no. 1, pp. 1–122, 2011.

R EFERENCES
[1] R.H. Lasseter, “MicroGrids”, in IEEE Power Engineering Society Winter
Meeting, 2002, vol. 1, pp. 305 – 308.
[2] R.H. Lasseter and P. Paigi, “Microgrid: a conceptual solution”, in
IEEE 35th Annual Power Electronics Specialists Conference, June 2004,
vol. 6, pp. 4285 – 4290. Gabriela Hug (S’05, M’08) was born in Baden,
[3] F. Katiraei, R. Iravani, N. Hatziargyriou, and A. Dimeas, “Microgrids Switzerland. She received the M.Sc. degree in elec-
management”, IEEE Power and Energy Magazine, vol. 6, no. 3, pp. 54 trical engineering from the Swiss Federal Institute
–65, June 2008. of Technology (ETH), Zurich, Switzerland, in 2004
[4] A.G. Tsikalakis and N.D. Hatziargyriou, “Centralized control for and the Ph.D. degree from the same institution in
optimizing microgrids operation”, IEEE Transactions on Energy Con- 2008. After her PhD, she worked in the Special
version, vol. 23, no. 1, pp. 241–248, 2008. Studies Group of Hydro One in Toronto, Canada and
[5] S. Conti, R. Nicolosi, S.A. Rizzo, and H.H. Zeineldin, “Optimal dis- since 2009 she is an Assistant Professor at Carnegie
patching of distributed generators and storage systems for MV islanded Mellon University in Pittsburgh, USA. Her research
microgrids”, IEEE Transactions on Power Delivery, vol. 27, no. 3, pp. is dedicated to control and optimization of electric
1243–1251, 2012. power systems.
[6] C.M. Colson and M.H. Nehrir, “Algorithms for distributed decision-
making for multi-agent microgrid power management”, in IEEE Power
and Energy Society General Meeting, 2011.
[7] A.L. Dimeas and N.D. Hatziargyriou, “A MAS architecture for micro-
grids control”, in 13th International Conference on Intellgient Systems
Application to Power Systems, Arlington, USA, 2005.
[8] S. Kar, J.M.F. Moura, and K. Ramanan, “Distributed parameter esti- Soummya Kar (S’05, M’10) received the B.Tech.
mation in sensor networks: nonlinear observation models and imperfect degree in Electronics and Electrical Communication
communication”, IEEE Transactions on Information Theory, vol. 58, Engineering from the Indian Institute of Technolo-
no. 6, pp. 3575 – 3605, June 2012. gy, Kharagpur, India, in May 2005 and the Ph.D.
[9] S. Kar and G. Hug, “Distributed robust economic dispatch in power degree in electrical and computer engineering from
systems: A consensus + innovations approach”, in IEEE Power and Carnegie Mellon University, Pittsburgh, PA, in 2010.
Energy Society General Meeting, July 2012. From June 2010 to May 2011 he was with the EE
[10] A.J. Conejo, F.J. Nogales, and F.J. Prieto, Decomposition Techniques in Department at Princeton University as a Postdoctoral
Mathematical Programming, Springer, 2006. Research Associate. He is currently an Assistant
[11] A.N. Venkat, I.A. Hiskens, J.B. Rawlings, and S.J. Wright, “Distributed Research Professor of ECE at Carnegie Mellon U-
MPC strategies with application to power system automatic generation niversity. His research interests include performance
contro”, IEEE Transactions on Control Systems Technology, vol. 16, no. analysis and inference in large-scale networked systems, adaptive stochastic
6, pp. 1192–1206, 2008. systems, stochastic approximation, and large deviations.
[12] S. Talukdar, J. Dong, P. Hines, and B. Krogh, “Distributed model
predictive control for the mitigation of cascading failures”, 2005, pp.
4440–4445.
 11

Chenye Wu (S’11, M’14) is a postdoctoral re-

searcher at Carnegie Mellon University, working
with Profs. Gabriela Hug and Soummya Kar. Before
joining CMU, he received the doctoral degree in
Computer Science from IIIS, Tsinghua University,
in July 2013, under the supervision of Prof. An-
drew Yao. He was the best paper co-recipients of
IEEE SmartGridComm 2012, and IEEE PES Gen-
eral Meeting 2013. Currently, he is working on the
optimization, distributed control and game-theoretic
analysis of power systems and electricity market.

A PPENDIX
Traditionally, demand functions are modeled as shown in
Fig. 18(a) with the load taking positive values. The marginal
demand function
dC  /dPL = an PL + bn (34)
with an < 0, bn > 0 is decreasing with increasing P L and has

positive lower and upper bounds P L and P L . The integral of
the marginal demand function provides the benefit function
shown in the lower part of Fig. 18(a) which is concave and
continuously increasing in the feasible range.
If we now define power to be negative when being con-
dC  /dP dC/dP
sumed and want to ensure that the marginal demand values
are equal to the ones from the traditional formulation, i.e.,
dCL /dPL = an PL + bn = dCL /dPL , (35)
the parameters in the new marginal demand function need to
fulfill an = −an > 0 and bn = bn > 0. The resulting benefit P L 
PL PL PL
function
C(P ) C(P )
1 1
CL (PL ) = an PL2 + bn PL = − an PL2 + bn PL = −C  (PL ) PL PL
2 2
(36)
which is also reflected in Fig. 18(b).
Hence, the two models are equivalent to each other where
the advantage of the employed model is that powers P n can P L 
PL
be summed and set equal to zero without having to make a (a) (b)
distinction between load or generation and the same holds for
Fig. 18. Equivalent load models.
the calculation of the social welfare.
The argumentation for the storage model is very similar. It
can be considered to be a load if the device is being charged,
i.e., Pn < 0, and it is equivalent to a generator if it is being
discharged, i.e., P n > 0. Consequently, with c n = 0, the
demand/cost function and the marginal demand/cost function
shown in Fig. 1(c) result.