Sei sulla pagina 1di 10

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
1

A Consensus-based Distributed Computational


Intelligence Technique for Real-Time Optimal
Control in Smart Distribution Grids
Kumar Utkarsh, Student Member, IEEE, Anupam Trivedi, Member, IEEE, Dipti Srinivasan, Senior
Member, IEEE, and Thomas Reindl

get an independent estimate about the global value of a desired


AbstractIn real-time large-scale optimization problems, such variable. As such, averaging consensus has been extensively
as in smart grids, centralized algorithms may face difficulties in used in literature for estimation [1].
handling fast-varying system conditions, such as high variability
of renewable-based distributed generators (DGs) and A. Applications of consensus-based distributed algorithms
controllable loads (CLs). Further, centralized algorithms may Consensus-based distributed optimization algorithms find
encounter computation and communication bottlenecks while
use in a wide variety of areas, such as in intelligent
handling large number of variables. To tackle these issues,
consensus-based distributed strategies have been proposed transportation systems [2], energy management in smart grids
recently. However, distributed computational intelligence (CI) [3], [4], synchronization of coupled oscillators [5], and control
based techniques can provide a much better near-optimal of multi-robot systems [6], among others. An important
solution within fewer iterations of the algorithm, which is a application of consensus-based distributed algorithms is in the
critical requirement in smart grids. Therefore, in this paper, a optimization of large-scale systems with a very high number
consensus-based dimension-distributed CI technique is proposed
of control variables. In this framework, consensus-based
for real-time optimal control in smart distribution grids, in which
large number of DGs and CLs are present. The proposed algorithms provide a platform in which the need for a
approach considers each DG or CL as a separate private entity, centralized optimizer is obviated and computational effort is
which is more relevant from the perspective of smart grid evenly spread across multiple entities.
optimization. In the proposed consensus-based framework, each
DG or CL is associated with an agent, and each agent is allowed B. Brief introduction to smart grids and the real-time optimal
to communicate only with its neighboring agents. The control problem
effectiveness of the proposed approach in terms of convergence, The future power system infrastructure will differ
adaptability and optimality with respect to a centralized
significantly from the present one, with increased integration
algorithm and a benchmark algorithm is shown through
simulations on 30-node and 119-node distribution test systems. of especially-renewable-based distributed generation sources
(DGs), variable controllable loads (CLs), and communication
Index TermsComputational intelligence, distributed control, and control capabilities. Thus, the large number of
particle swarm optimization, reactive power control, smart grids. controllable equipment present in the future power system will
result in increased degrees of freedom, and can help to further
I. INTRODUCTION optimize the desired objectives. This future power system,
appropriately named as a smart grid, will utilize the
C ONSENSUS based optimization forms a very important
part of distributed computing literature. In consensus
problems, the primary objective is for a group of individuals
aforementioned resources to make electricity generation,
transmission, and distribution more reliable, secure, and as
with varying opinions to reach a common understanding. efficient as possible [7].
Consensus is especially important in situations where However, the ever increasing penetration of DGs and CLs
individuals value privacy, and are reluctant to share their local in smart grids will introduce such scalability and variability
information with a centralized entity. In networked systems, issues which the existing centralized control architectures may
consensus plays a key role in enabling individual entities to be incapable of handling [8], [9]. Secondly, there will be
privacy concerns, wherein DG and CL owners will be
reluctant to share their local information to a centralized entity
This work was supported by the National Research Foundation under [7], [10]. Therefore, recently, there has been an increasing
Grant R-712-000-032-281. focus on developing distributed algorithms to tackle such
K. Utkarsh, A. Trivedi, and D. Srinivasan are with the Department of
issues. Further, in real-time optimal control in smart grids,
Electrical and Computer Engineering, National University of Singapore,
Singapore 117576 (e-mail: utkarsh.k@u.nus.edu; eleatr@nus.edu.sg; power loss minimization and voltage regulation are two
dipti@nus.edu.sg). T. Reindl is with the Solar Energy Research Institute of important problems that the system operator needs to solve
Singapore, National University of Singapore, Singapore 117574 (e-mail: [8], [9], [11], [12]. In a power distribution system, power loss
thomas.reindl@nus.edu.sg)
minimization can be viewed typically as the cost minimization

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
2

aspect, because a lower power loss results in a lower (i.e. control variables) of the solution vector. Thus, dimension-
electricity purchase from the external grid. On the other hand, distributed EAs are truly distributed implementations with
voltage regulation is the technical aspect, in which voltages respect to a practical problem, in which each dimension is
are to be typically regulated within a certain range around the controlled by a separate controller or agent.
nominal value, to maintain a reliable electricity supply to the Existing works on dimension-distributed EAs can be further
loads. Therefore, in this paper, our focus is to solve these classified [17] as utilizing either coevolution models or game-
problems in a real-time and distributed manner. theoretic multi-agent models. The game-theoretic multi-agent
model requires global objectives to be expressed as a sum of
C. Existing consensus-based distributed techniques for real-
local objectives of agents, which may not always be possible
time optimal control in smart grids
[18]-[20]. In the coevolution model, on the other hand, an
Several authors have recently proposed consensus-based agent needs to know the variables of other agents as well,
distributed strategies to tackle the real-time optimization which may lead to violation of privacy [21]. Also, an agent
problem in smart grids considering reactive power availability performs several generations of evolution of its primary
of DGs. In [11], power loss is assumed to be a function of variable (corresponding to its own dimension), while
voltage magnitudes and the DGs reach a common utilization considering secondary variables (corresponding to other
ratio to minimize voltage deviations. In [8], [13], a dual- agents dimensions) to be clamped. This results in the
decomposition based feedback algorithm to minimize power evolution being performed with respect to outdated secondary
losses with voltage constraints is presented considering X/R variables. Then, in the communication phase, each agent
ratio of power lines to be the same. In [3], power loss based on updates its secondary variables, which need to be broadcast by
general loss coefficients is minimized using an equal- other agents, resulting in extra communication time overhead
incremental criterion based algorithm. In [14], distributed [22], [23].
dual-decomposition based algorithm is developed to minimize
operation cost of DGs, however, system power loss is ignored. E. Proposed approach and Contributions
In [9], a distributed sub-gradient based algorithm is developed The literature survey presented above indicates that real-
under several assumptions to control voltages and minimize time optimization involving distributed EAs is an emerging
power losses. While, in [15], [16], voltage control is achieved topic in computational intelligence. Thus, in this paper, we
using proportional reactive power sharing while considering propose a consensus-based dimension-distributed CI technique
power lines to be lossless. The literature survey on real-time for real-time optimal control in smart distribution grids. The
optimal control in smart grids shows that the existing major contributions of this work can be summarized as
consensus-based distributed algorithms are developed under follows:
several assumptions, which may not hold true in realistic 1) It proposes a consensus-based dimension-distributed CI
power networks. algorithm for optimization in a large-scale system, using
only realistic assumptions.
D. Distributed computational intelligence techniques for
high-dimensional problems 2) In this work, an agent exchanges estimated states of the
system with other agents, rather than the control variables.
In comparison to the existing consensus-based distributed This results in the privacy of individual agents being
optimization techniques, distributed computational maintained, which is particularly important in smart grid
intelligence (CI) based techniques can provide a near-optimal scenarios.
solution for high-dimensional problems within fewer iterations 3) In comparison to existing works on dimension-distributed
of the algorithm [17], which is a critical requirement in smart EAs, the proposed algorithm ensures minimum
grids with a high penetration of renewable DGs and CLs. communication overhead, while being able to handle any
Therefore, as a part of distributed CI techniques to deal with generalized global objective function.
such problems of high dimensions, distributed evolutionary 4) To the best of our knowledge, the proposed consensus-
algorithms (EAs) have been developed. based dimension-distributed CI algorithm has not been
Distributed EAs have been further classified as population- studied in the context of real-time optimal control in
distributed or dimension-distributed [17]. Population- smart grids so far.
distributed EAs assume each agent to consist of a probable
The rest of the paper is organized as follows. Section II
solution vector, say a particle in particle swarm optimization
explains the real-time optimal control problem in smart grids;
(PSO) or a chromosome in genetic algorithm (GA), i.e.
Section III provides details about the proposed fully
where D is the number of control variables involved in
distributed CI algorithm for real-time optimal control in smart
optimization. Therefore, in population-distributed EAs, each
grids; Section IV provides simulation studies for the proposed
agent needs information about all the D control variables.
approach; and finally Section V concludes the paper.
However, in a real scenario, the D control variables may
belong to separate private entities and, therefore, it is much
II. PROBLEM FORMULATION
more reasonable to associate an agent with each of the D
control variables. This is the rationale of dimension- In this section, the problem formulation for real-time
distributed EAs, in which agents correspond to the dimensions optimal control of DGs and CLs for power loss minimization
and voltage regulation is presented, before which a description

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
3

of the envisioned system is provided. where and are the minimum and maximum allowed
operating voltages.
2) Reactive power constraints: The reactive power
availability of DG or CL converters is constrained.
( []) (, ) (4)
where , is calculated according to (1), and z is the present
iteration.
3) Reactive power ramp-rate constraints: Reactive power
change between two consecutive iterations needs to be
constrained.
Fig. 1. Physical power system layer (power lines shown in blue) with DGs [ 1] [] [ 1] + (5)
and loads; and cyber layer with agents and communication infrastructure where is the ramp-limit of converter i.
(shown in black dotted lines).
A hypothetical section of a smart distribution grid with C. Relation between control variables and states
several DGs and loads is shown in Fig. 1, in which power In this work, we have considered reactive power (i.e. ) of
lines are overlaid with communication infrastructure and each DG and CL converters as control variables. However, in
node is being monitored and controlled by an agent. It is noted power systems, phasorial voltages, i.e. , of each node are
that the DGs are coupled to the grid usually through power considered as states of the system. They are related according
electronic converters, which are designed to provide fast to the following equation.
response and efficiently transmit energy from the DGs to the
grid. These converters are designed for maximum active = ( ) (6)
=1
power availability of the DGs, and since DGs normally
It is clear from (6) that the control variables and the state
operate at a capacity lower than their converters rating, the
variables have a non-linear relationship, and thus we need to
remaining capacity of the converters can be utilized to provide
make use of the following proposition, which provides an
reactive power support to the grid. The same is true for CLs,
approximate relation between node voltages and
where their converters can be utilized to provide reactive
active/reactive power injections at the nodes.
power support to the grid. Therefore, our aim in this paper is
Proposition: Node voltages relate to active and reactive
to utilize the reactive power injection/absorption capability of
power injections/absorptions at the nodes as follows:
such controllable equipment (i.e. DG and CL converters) to
efficiently achieve the overall objectives. 2 [(2: , 2: )]1 2 2
[ ] = 0 1 + [ ] (7)
Thus, the control variables in our problem are reactive 0
power injections/absorptions of power converters of DGs and

CLs, which are located at different nodes in the power system. where the phasorial voltage at node i, 1 = [1, ,1] , Y is
Further, we assume that each agent on a node can the admittance matrix, and 0 is the nominal voltage at the
communicate only with its neighboring agents (i.e. agents on slack node 1.
nodes connected physically with power lines), as shown in Proof: Firstly, it is noted that, in a power system, if
Fig. 1. Now, if the ith DG or CL is supplying/receiving an active/reactive power injections/absorptions were zero at all
active power to/from the grid, and its power converter rating nodes, the voltage at all the nodes would be equal to that of
is , , then the reactive power availability of its power the slack node [24]. Now, integrating the differential grid
converter, , , can be defined as follows: model in [25, Section II], with initial conditions as
2
aforementioned, (7) can be obtained.
2
, = (, 2 ) (1)
D. Decomposition method to increase computation speed
A. Objective function In a properly designed power system, a small change in the
In this paper, the objective function is to minimize active power injection/absorption at a particular node does not affect
power losses in the power system using optimal reactive the entire system equally. Thus, for a particular node i, the
power injection/absorption from DG and CL converters: whole power system can be divided into two parts-the set of
nodes which are strongly affected (set- ), and the other set of
(2) those nodes which are very weakly affected (set- ). These
=1 =1
sets of nodes can be determined using the -based
where is the number of nodes in the power system, is the decomposition method on the matrix [26]. Therefore, to
voltage magnitude at node i, and and are conductance enhance computation speed of the proposed algorithm, we put
and angle difference, respectively, between nodes i and k. a constraint such that the active/reactive power
B. Constraints injection/absorption at a node i only affects nodes in set- ,
1) Voltage constraints: Voltage magnitude at each node i but not the nodes in set- .
needs to be in limits.
< < (3)

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
4

III. PROPOSED CONSENSUS-BASED DIMENSION-DISTRIBUTED Further, the elements of state vector y may be dependent on
CI ALGORITHM FOR REAL-TIME OPTIMAL CONTROL OF SMART one or more elements of vector x, represented appropriately as
GRIDS follows:
1 1
In this section, the proposed consensus-based distributed CI
[ ] = [ ]
algorithm for real-time optimal control of smart grids is
presented in the context of a generic large-scale global
optimization problem. Firstly, consensus algorithm is = (: , ) (13)
=1
discussed, which is used by the agents to estimate the states of where is a matrix relating control variables x to states y. It is
the system. Secondly, a framework for a generalized large noted that (13) is a linearized form of relationship, which
scale optimization problem is discussed, and finally, the could be derived using suitable approximations. This relation
proposed algorithm is presented. is equivalent to the relation (7) in Section-II, with the state
A. Consensus Algorithm vector y corresponding to voltage vector V, and control
variable vector x corresponding to reactive power vector Q.
Consider a connected graph G with N nodes and E edges, in In this paper, we have implemented a dimension-distributed
which each node is assumed to be controlled by an agent. Any
implementation of particle swarm optimization (PSO), which
two agents are considered neighbors iff they have a common
is an instantiation of a CI algorithm. It is noted that the steps
edge , and each agent can communicate with its
involved for implementation of the proposed distributed PSO-
neighboring agents only. Now, let [] denote the
estimate of a parameter maintained by agent k at an iteration z. based algorithm can be translated into equivalent steps for a
The value may represent any quantity that the agents are similar population-based CI algorithm as well.
trying to reach a consensus on. We say that a network of N C. Brief description of the standard PSO algorithm
agents has reached consensus iff | | , , =
PSO [27] is a direct search and stochastic optimization
(1 ), , where is a very small positive number.
technique that is based on the behavior of a social group of
According to the consensus algorithm, each agent k updates its
estimates using the following discrete-time model: animals. In PSO, the decision of an ith particle , where
N is the dimension of the solution hyperspace, is influenced by
[ + 1] = , [] (8) three factors: the particles own velocity, the best known
=1
solution found by the particle, and the best known solution
where , represents the communication coefficient between found among all the particles. The basic equations that the
agents k and l. It is noted that , is non-zero only if the two standard PSO uses for a particle i are as follows:
agents are neighbors. Thus, the combined consensus algorithm [] = . [ 1] + 1 . 1. ( [ 1])
can be written as: +2 . 2. ( [ 1]) (14)
[ + 1] = [] (9) [] = [ 1] + [] (15)
where C is a matrix composed of elements , , , = where z is the present iteration, and are velocity and
(1 ), and is the column vector containing all . position, respectively, of particle i, w is the inertia factor, 1
According to the proof in [1], if the matrix C is doubly and 2 are two positive numbers, rand1 and rand2 are two
stochastic, i.e. all its rows and columns individually sum up to random numbers uniformly distributed in the range [0,1],
one, and if the eigenvalues of C are less than or equal to 1, is the best solution found so far by particle i, and
then all the will converge to a common estimate as given is the best solution found among all the particles. PSO
below: and its variants have been shown to efficiently solve large
1 scale non-linear problems while not suffering from the curse
[] = [0] , = (1 ) (10)
=1 of dimensionality or slow convergence [28]. Therefore,
The matrix C can be made to fulfil the aforementioned inspired by its properties, we will next present the proposed
conditions by using the metropolis algorithm [1]. consensus-based dimension-distributed PSO algorithm.

B. Framework for a generalized large scale optimization D. Proposed consensus-based dimension-distributed PSO-
problem based algorithm
Now, let us suppose that the generalized global objective In the proposed algorithm, N agents are considered
function to be minimized is as follows: corresponding to N control variables in a system, with each
min () (11)
agent i corresponding to a control variable (i.e. reactive
power output of a DG or CL converter in the formulated
where N is the dimension of the control variable vector, real-time optimal smart grid control problem). Each agent also
. Although the global objective function f(x) is generalized maintains P probable values of its control variable in the form
enough to capture all possible functions, it is more convenient of a column vector. Hereafter, we will refer to each such value
to write the function f in terms of a state vector, say vector of the vector as an element. Now, for each of its P elements,
, of dimension M. an agent has to estimate its contribution towards the state
min () (12) vector y of the system using (13). Specifically, an agent i, for

each element {1, , }, will estimate its contribution as

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
5


(: , ) and will share this estimate with neighboring agents Step-2 (Synchronization): The agent then communicates with
using consensus algorithm to get an updated estimate of state its neighboring agents and updates its state estimates as
vector y. In the context of the formulated real-time optimal follows:
smart grid control problem, each agent i, for each element p,

will estimate its contribution towards the voltage vector V in ,
1 [0] , (1, ) [0]
=1
the system using

(: , ) and will share this estimate with [ ]= (18)
,
[0]
neighboring agents using consensus algorithm to get an , (, ) [0]
[ =1 ]
updated estimate of the voltage vector V. This framework is
also depicted in Fig. 2, where the element vectors and the The agent now calculates the objective function f as in (12)
estimated state vectors of two agents i and j are shown. (which is equivalent to power loss in the formulated smart grid
problem (2)) for each element p. It then finds out the element-
best value for each element p (i.e. the value of the control
variable corresponding to the best known objective
function found so far by the element p) and the global-best

value (i.e. the value of the control variable corresponding to
the best known objective function found so far among all the P
elements). It is noted that the 1-dimensional element-best and
global-best values of an agent correspond to the N-
dimensional particle-best and global-particle-best values in
the standard PSO. The velocities of the elements are also
randomly initialized in this step. We now move to the main
Fig. 2. Framework of agents for the proposed distributed algorithm. part of the algorithm.
Step-3 (Evolution of elements): Consider this step as the start
It is noted that the real-time optimal control problem in of iteration z. To eliminate the phenomenon of stagnation in
smart grids is a dynamic optimization problem, in the sense the proposed fully distributed algorithm, two important
that the system parameters (i.e. active power modifications are made:
generation/consumption of DGs and CLs) vary with time. 1) Velocity and position (i.e. element value) update
Therefore, let us consider a control-horizon , which is equations of the global-best element are integrated with
divided into H control-cycles, and assume that the system
the Guaranteed Convergence PSO (GCPSO) [29], [30]. In
parameters are varying across these control-cycles. Further,
GCPSO, the modified velocity update equation of the
we consider that in each such control-cycle {1, , }, the
global-best element is as follows:
system remains static (i.e. the system parameters or the active
power generation/consumption of DGs and CLs remain , [] = . , [ 1] , [ 1] + , + [] (19)
constant), and the agents perform several iterations of the where , and , are velocity and position (i.e. element
proposed algorithm until convergence to optimize the system value), respectively, of the global-best element,
for this control-cycle before moving on to the next control- and , is its element-best value. w is a weighting
cycle + 1. Considering the dynamics in the system, we will
now present the execution steps of the algorithm to be coefficient, r is a random number sampled from the
performed by each agent i in the control-horizon . interval (-1,1), and () is a scaling factor determined as
Step-0 (Commencement of control-horizon ): Let us follows:
consider this step as the starting point of the control- [0] = 1
horizon , i.e. this step is the beginning of control-cycle = 2[], >
1. [ + 1] = { 0.5[], > (20)
Step-1 (Initialization): The agent i initializes a column vector [],
of its probable values, according to the following: where and are thresholds. The thresholds are such
[0] = , + 2 , (, 1) (16) defined that whenever the global-best element improves
where = [1 , , ] is a P-element column vector, , its element-best value , , the success count is
is the maximum allowed value of the control variable , and incremented and the failure count is set to 0, and vice
(. ) generates a column vector of random numbers versa. However, the success and failure counts are both
between 0 and 1. It is noted that is analogous to the reactive set to 0 whenever the global-best element itself changes.
power of DG and CL converters. For each element p in the These modifications ensure that the global-best element
column vector obtained in Step 1, the agent then initializes its keeps performing a random search around its position,
estimates of the state vector (or equivalently the voltage vector resulting in its escaping from a local minimum.
V) as follows: 2) The velocities of all elements, except the global-best
,
1 [0] element, are checked if they have fallen below a certain

[ ] = (: , ) [0] (17) threshold. If yes, then they are reset to random values in
, the allowable range. This allows the elements to keep
[0]

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
6

looking for better solutions, while at the same time of its control variable (i.e. ) in the following manner. Firstly,
retaining knowledge learnt by the global-best element. each agent randomly selects 10% of its elements and replaces
Step-4 (Adaptation and Synchronization): For each element them with random immigrants, which basically are randomly
p, the agent now estimates the new state vector, due to change chosen values in the allowable range. Secondly, each agent
in the elements value at the present iteration, and randomly selects 40% of the elements and replaces them with
communicates the state vector to its neighboring agents as elite immigrants, which are random mutations within +20% of
follows: the previously obtained global-best element value. After these
, ,
modifications, the control is again transferred to Step-4.
1 [] 1 [ 1]
Remark 1: The proposed algorithm is fully distributed, in
[ ]=[ ] + (: , ) ( [] [ 1]) (21)
, , the sense that it does not require any high-level or external
[] [ 1]
controller to coordinate the actions of the agents. The agents
,
,
1 [] , 1 [] only need to store the matrices C and in order to do all the
=1
[ ]= (22) calculations; and the computational burden on the agents, at
,
[] ,
, [] each iteration z, is also small.
[ =1 ] Remark 2: In the proposed algorithm, privacy of individual
The agent now calculates the objective function f as in (12), agents is maintained, in that a particular agent, with its
for each element p, and updates the element-best and global- available information, cannot determine converged global-best
best values. element values (i.e. reactive power values) of its neighboring
Step-5 (Real-time implementation of obtained solution): agents. This is explained as follows. Referring to (18) and (22)
Since the proposed algorithm is a real-time algorithm, each for an element p, we can see that an agent j is exchanging only
agent immediately applies its global-best element value found its estimated state vector at each iteration z, rather than its

at each iteration to its respective node. This strategy ensures elements actual value (i.e. a probable reactive power
that the agents respond as fast as possible to changing system value). Now, one can argue from (17) and (21) that an
conditions, and do not wait until convergence is achieved to estimated state vector is a function of the elements present
apply solutions. and past values. Further, it can also be argued that since all
Step-6 (Check for convergence): The algorithm is assumed agents know matrices C and , any particular agent i can find
to have converged if the agents ascertain that the change in the final global-best element value of its neighboring agent j
their global-best element values is below a threshold for a (by simply tracking agent js state estimates until
certain number of consecutive iterations, or if a maximum convergence). However, it must be noted that the agent i does
number of iterations is reached. If the algorithm has not not know what the indices of its neighboring agents are, and
converged, the control is transferred to Step-3 again, and the
therefore does not exactly know the column vector (:,j) for
iteration counter z is also incremented to z+1. Further, if the
its neighboring agent with index j. Therefore, agent i cannot
algorithm has converged and the present control-cycle has
find out the element values of agent j, even though it knows
ended, it is incremented to + 1 and the control is transferred
to Step-7. the state vectors for agent j. Thus, privacy of any individual
Step-7 (Population update for dynamic optimization): It agent is maintained by the proposed algorithm.
has previously been noted in this sub-section that the real-time
optimal control problem in smart grids is a dynamic IV. SIMULATION RESULTS AND DISCUSSION
optimization problem, in which the system parameters vary The proposed algorithm is implemented using MATLAB
with time. This variation manifests itself in the objective software on two distribution test systems: 1) the standard
function f to change with each control-cycle, which results in IEEE 30-node system, which is a meshed network, and which
the optimal solution vector for a particular control-cycle, to is modified to simulate a distribution system with nominal
vary with each control-cycle, where is defined as follows. voltage 11 kV [25]; and 2) an 11 kV radial distribution system
= argmin () (23) with 119 nodes [32], which effectively simulates a large-scale

To deal with such dynamics, an obvious approach could be system to test scalability of the proposed algorithm. The
to restart the entire algorithm for each control-cycle. However, voltage variations at all the nodes are limited to +3% of the
this restart-based algorithm would be very inefficient, because nominal value. For both the systems, DGs and CLs are
we essentially lose all the knowledge gained from the previous assumed to have been placed after proper siting and sizing
control-cycle, which could have otherwise been utilized to studies, and have converters of rating 5 MVA each.
reach the new optimal solution faster [31]. Based on this logic In this paper, two case studies each were conducted for both
and to make the algorithm more flexible, evolutionary 30-node and 119-node systems to demonstrate the efficacy of
dynamic optimization (EDO) technique is integrated in the the proposed distributed algorithm:
proposed algorithm. The integrated EDO-based approach is 1) Case Study 1 convergence and adaptability of the
able to effectively manage system dynamics, in that a fast proposed algorithm, under dynamic renewable outputs and
convergence to the optimal solution is achieved after a change loads for a control-horizon = 10, is investigated and
in system conditions [31]. benchmarked with a reference consensus-based distributed
In this integrated EDO-based approach, at the end of each algorithm [11].
control-cycle, each agent i modifies the final converged vector

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
7

2) Case Study 2 optimality of the proposed algorithm is


investigated with respect to a centralized controller based
algorithm for a control-horizon = 10.
A. Simulation studies on 30-node test system
In this section, we investigate the performance of the
proposed algorithm on the 30-node distribution system, and
since the number of nodes are relatively low, there was no
need to limit the communication of a node (agent) to only its
strongly coupled nodes. The DG penetration is set to 65.2%,
with 12 DGs and 8 CLs in the system. Thus, the number of
control variables involved is 20.
1) Case study 1 Fig. 3. Power loss plots for control-horizon = 10 for 30-node system.
In this case study, we investigate convergence and
adaptability of the proposed algorithm to changing conditions,
i.e. variability of renewables and loads, for the given control-
horizon = 10. Further, comparisons are drawn among 1) the
EDO-based proposed approach, 2) the restart-based proposed
approach (in which after every control-cycle, each agent
reinitializes its elements values and velocities for
optimization of the next control-step), and 3) the reference
consensus-based distributed algorithm in [11].
The results of the simulation study are shown in Fig. 3, 4
and 5. Fig. 3 displays power loss progression for the
aforementioned three algorithms for the control-horizon =
10, and Fig. 4 provides a zoomed-in view of the 2nd control-
cycle for visual clarity. It can be observed from both the Fig. 4. Enlarged view of the 2nd control-cycle plots for 30-node system.
figures that the EDO-based algorithm is able to find better
solutions and converge faster than the other two algorithms.
This property can be appropriately quantified in terms of mean
best-of-generation value [33]:
1 1
= ( , ) (24)
=1 =1

where G is the number of generations (i.e. iterations), R is the


number of runs of the algorithm, and , is the global-best
power loss value obtained at ith iteration and jth run. Thus, the
lower the , the better the algorithm is in finding good
solutions in a shorter time. Referring to Table I, we can see
that the EDO-based proposed algorithm has the lowest
Fig. 5. Node voltage plots for 30-node system for EDO-based algorithm.
value, thus outperforming the other two algorithms. Further,
Fig. 5 shows that the EDO-based algorithm is able to
effectively maintain node voltages inside the prescribed limits
of + 3% in a fully distributed manner.
2) Case study 2
In this case study, we investigate the optimality of the
solution reached by the EDO-based algorithm with respect to
the centralized controller based algorithm. The results are
presented in Fig. 6, which shows average converged values for
20 runs of both the algorithms. It can be seen that the EDO-
based algorithm is able to find a solution almost similar to the
centralized algorithm. In fact, the average error is found to be
0.94% for the given control-horizon = 10, which indicates
convergence to near-optimality by the EDO-based proposed Fig. 6. Converged power loss values for Case Study 2 for 30-node system.
algorithm.

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
8

TABLE I better solutions and converge faster than the other two
MEAN BEST-OF-GENERATION VALUE ( ) FOR R=20
algorithms. Also, referring to Table I, we can see that the
EDO-based Restart-based Reference
System algorithm algorithm
EDO-based proposed algorithm has the lowest value,
algorithm
30-
thus outperforming the other two algorithms. Further, Fig. 10
1.162 1.228 1.231
node shows that the EDO-based algorithm is able to effectively
119- 0.975 1.012 1.135 maintain node voltages inside the prescribed limits of + 3% in
node a fully distributed manner, even for a large 119-node system.
It is noted from Fig. 8 that the reference algorithm is not
B. Simulation studies on 119-node test system performing well, especially for control-cycles 5 to 10. This is
In this section, we investigate scalability aspect of the due to some of its assumptions, such as not considering phase
proposed algorithm by conducting simulation studies on a angles of the nodes, which are causing it to accumulate a large
large 119-node distribution test system. The DG penetration is error for the large 119-node system.
set to 57.6% in the following studies, with 48 DGs and 32 CLs 2) Case study 2
in the system. Thus, the number of control variables involved In this case study, we investigate the optimality of the
is 80, which represents a large system effectively. It is noted solution reached by the EDO-based algorithm with the
that consensus on a large-scale system may take a bit longer to centralized controller based algorithm.
be achieved, especially if the network graph is sparsely
connected. This may result in delays in the convergence of the
algorithm. Therefore, as discussed in Section-II, we can divide
the complete power system into two sets and for
each node i. For example, Fig. 7 shows the sensitivity values
(magnitude of ) of nodes 27 and 73. Therefore, the effect of
node 27, for example, is assumed to be limited to the nodes in
set-27 only. This implies that even a very-large system can be
broken down into smaller sub-systems for estimating the
effect of a node i, and, therefore, enables distributed
optimization of such large systems effectively. However, it
must be noted that too much restriction on this set of nodes
may lead to a solution which is much farther from the optimal
solution, despite having converged faster. Fig. 8. Power loss plots for control-horizon = 10 for 119-node system.

Fig. 7. Sensitivity values of nodes 27 and 73, for the 119-node system. Fig. 9. Enlarged view of the 2nd control-cycle power loss plots for 119-node
system.
1) Case study 1
In this case study, we investigate convergence and
adaptability of the proposed algorithm to changing conditions,
i.e. variability of renewables and loads, for the given control-
horizon = 10. As for the 30-node system, comparisons are
drawn here among 1) the EDO-based proposed approach, 2)
the restart-based proposed approach, and 3) the reference
consensus-based distributed algorithm in [11].
The results of the simulation study are shown in Fig. 8, 9
and 10. Fig. 8 displays power loss progression for the
aforementioned three algorithms for the given control-
horizon = 10, and Fig. 9 provides a zoomed-in view of the
2nd control-cycle for visual clarity. It can be observed from
both the figures that the EDO-based algorithm is able to find Fig. 10. Node voltage plots for 119-node system for EDO-based algorithm.

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
9

[4] Y. Xu, Z. Yang, W. Gu, M. Li, and Z. Deng, Robust Real-Time


Distributed Optimal Control Based Energy Management in a Smart
Grid, IEEE Transactions on Smart Grid, vol. PP, no. 99, pp. 1-12,
2015.
[5] V. M. Preciado and G. C. Verghese, "Synchronization in Generalized
Erds-Rnyi Networks of Nonlinear Oscillators," in Proceedings of the
44th IEEE Conference on Decision and Control, 2005, pp. 4628-4633.
[6] X. Xiaorui and E. H. Abed, "Formation Control with Virtual Leaders
and Reduced Communications," in Proceedings of the 44th IEEE
Conference on Decision and Control, 2005, pp. 1854-1860.
[7] M. Yazdanian and A. Mehrizi-Sani, "Distributed Control Techniques in
Microgrids," IEEE Transactions on Smart Grid, vol. 5, pp. 2901-2909,
2014.
[8] S. Bolognani, R. Carli, G. Cavraro, and S. Zampieri, "Distributed
Reactive Power Feedback Control for Voltage Regulation and Loss
Fig. 11. Converged power loss values for Case Study 2 for 119-node system.
Minimization," IEEE Transactions on Automatic Control, vol. 60, pp.
The results are presented in Fig. 11, which shows average 966-981, 2015.
converged values for 20 runs of both the algorithms. It can be [9] W. Zhang, W. Liu, X. Wang, L. Liu, and F. Ferrese, "Distributed
Multiple Agent System Based Online Optimal Reactive Power Control
seen that the EDO-based algorithm is able to find a solution
for Smart Grids," IEEE Transactions on Smart Grid, vol. 5, pp. 2421-
almost similar to the centralized algorithm, and the average 2431, 2014.
error is found to be 2.31% for the given control-horizon = [10] W. Shi, X. Xie, C. C. Chu, and R. Gadh, "Distributed Optimal Energy
10. It is noted that this increase in error percentage, as Management in Microgrids," IEEE Transactions on Smart Grid, vol. 6,
compared to the 30-node system, can be attributed to the pp. 1137-1146, 2015.
[11] A. Maknouninejad and Z. Qu, "Realizing Unified Microgrid Voltage
tradeoff of using the -based decomposition method. Profile and Loss Minimization: A Cooperative Distributed Optimization
and Control Approach," IEEE Transactions on Smart Grid, vol. 5, pp.
V. CONCLUSION 1621-1630, 2014.
[12] L. Yu, D. Czarkowski, and F. d. Leon, "Optimal Distributed Voltage
In this paper, a consensus-based dimension-distributed CI Regulation for Secondary Networks With DGs," IEEE Transactions on
technique is proposed for real-time optimal control in smart Smart Grid, vol. 3, pp. 959-967, 2012.
distribution grids, in which large number of DGs and CLs are [13] S. Bolognani and S. Zampieri, "A Distributed Control Strategy for
present. In the proposed algorithm, each control variable is Reactive Power Compensation in Smart Microgrids," IEEE Transactions
considered as a separate private entity, which is relevant from on Automatic Control, vol. 58, pp. 2818-2833, 2013.
[14] P. Srikantha and D. Kundur, "Distributed Optimization of Dispatch in
the perspective of practical applications, such as in smart grid Sustainable Generation Systems via Dual Decomposition," IEEE
optimization. In comparison to existing consensus-based Transactions on Smart Grid, vol. 6, pp. 2501-2509, 2015.
distributed techniques for smart grid optimization, the [15] J. Schiffer, T. Seel, J. Raisch, and T. Sezi, "Voltage Stability and
proposed distributed CI technique can provide a near-optimal Reactive Power Sharing in Inverter-Based Microgrids With Consensus-
Based Distributed Voltage Control," IEEE Transactions on Control
solution within fewer iterations of the algorithm, which is a
Systems Technology, vol. 24, pp. 96-109, 2016.
critical requirement in smart grids with a high penetration of [16] D. He, D. Shi, and R. Sharma, "Consensus-based distributed cooperative
renewable DGs and CLs. The efficacy of the proposed control for microgrid voltage regulation and reactive power sharing," in
algorithm was shown through simulations on 30-node and IEEE PES Innovative Smart Grid Technologies, Europe, 2014, pp. 1-6.
119-node distribution test systems vis--vis convergence, [17] Y.-J. Gong, W.-N. Chen, Z.-H. Zhan, J. Zhang, Y. Li, Q. Zhang, et al.,
"Distributed evolutionary algorithms and their models: A survey of the
adaptability and optimality. Also, the proposed distributed CI state-of-the-art," Applied Soft Computing, vol. 34, pp. 286-300, 9// 2015.
algorithm was benchmarked to prove effectivenss over a [18] F. Seredynski, A. Y. Zomaya, and P. Bouvry, "Function Optimization
recently published consensus-based distributed algorithm, and with Coevolutionary Algorithms," in Intelligent Information Processing
a centralized controller based algorithm. Further, we note that and Web Mining: Proceedings of the International IIS: IIPWM03
since the proposed algorithm has been developed for a Conference held in Zakopane, Poland, June 25, 2003, M. A. Kopotek,
S. T. Wierzcho, and K. Trojanowski, Eds., ed Berlin, Heidelberg:
generalized large-scale optimization problem, it can be Springer Berlin Heidelberg, 2003, pp. 13-22.
extended beyond smart grid optimization to other challenging [19] M. K. Jalloul and M. A. Al-Alaoui, "A distributed Particle Swarm
real-time non-linear global optimization problems as well. Optimization algorithm for block motion estimation using the strategies
of diffusion adaptation," in Signals, Circuits and Systems (ISSCS), 2015
International Symposium on, 2015, pp. 1-4.
REFERENCES
[20] Y. Wakasa and S. Nakaya, "Distributed particle swarm optimization
[1] R. Olfati-Saber, J. A. Fax, and R. M. Murray, Consensus and using an average consensus algorithm," in 2015 54th IEEE Conference
Cooperation in Networked Multi-Agent Systems, Proceedings of the on Decision and Control (CDC), 2015, pp. 2661-2666.
IEEE, vol. 95, no. 1, pp. 215-233, 2007. [21] R. Subbu and A. C. Sanderson, "Modeling and convergence analysis of
[2] M. Saeednia, and M. Menendez, A Consensus-Based Algorithm for distributed coevolutionary algorithms," IEEE Transactions on Systems,
Truck Platooning, IEEE Transactions on Intelligent Transportation Man, and Cybernetics, Part B (Cybernetics), vol. 34, pp. 806-822, 2004.
Systems, vol. PP, no. 99, pp. 1-12, 2016. [22] Z. Jun, H. S. H. Chung, and W. L. Lo, "Pseudocoevolutionary genetic
algorithms for power electronic circuits optimization," IEEE
[3] C. Zhao, J. He, P. Cheng, and J. Chen, Consensus-Based Energy
Transactions on Systems, Man, and Cybernetics, Part C (Applications
Management in Smart Grid With Transmission Losses and Directed
and Reviews), vol. 36, pp. 590-598, 2006.
Communication, IEEE Transactions on Smart Grid, vol. PP, no. 99, pp.
1-13, 2016.

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TETCI.2016.2635130, IEEE
Transactions on Emerging Topics in Computational Intelligence
10

[23] K. C. Tan, Y. J. Yang, and C. K. Goh, "A distributed Cooperative Thomas Reindl received the Ph.D. degree in
coevolutionary algorithm for multiobjective optimization," IEEE natural sciences from the University of Regensburg,
Transactions on Evolutionary Computation, vol. 10, pp. 527-549, 2006. Germany, in 1996 and the MBA degree from
[24] "Load representation for dynamic performance analysis [of power INSEAD, Singapore, in 2001. He was with
systems]," IEEE Transactions on Power Systems, vol. 8, pp. 472-482, SIEMENS Corporate R&D Labs, Munich,
1993. Germany. From 1998 to 2006, he was a Chief
Operating Officer with ILIOTEC, Weiden,
[25] K. Utkarsh, D. Srinivasan and T. Reindl, "Distributed control of
Germany, and was later the Managing Director. In
renewable generators with energy storage," 2016 IEEE Power & Energy
2010, he was a Group Leader with Photovoltaic
Society General Meeting, Boston, MA, 2016 (Accepted). Systems Technology, Solar Energy Research
[26] M. Amano, A. I. Zecevic, and D. D. Siljak, "An improved block-parallel Institute, Stuttgart, Germany, where he currently holds the position of Deputy
Newton method via epsilon decompositions for load-flow calculations," Chief Executive Officer.
IEEE Transactions on Power Systems, vol. 11, pp. 1519-1527, 1996.
[27] J. Kennedy and R. Eberhart, "Particle swarm optimization," in Neural
Networks, 1995. Proceedings., IEEE International Conference on, 1995,
pp. 1942-1948 vol.4.
[28] Y. d. Valle, G. K. Venayagamoorthy, S. Mohagheghi, J. C. Hernandez,
and R. G. Harley, "Particle Swarm Optimization: Basic Concepts,
Variants and Applications in Power Systems," IEEE Transactions on
Evolutionary Computation, vol. 12, pp. 171-195, 2008.
[29] F. v. d. Bergh and A. P. Engelbrecht, "A new locally convergent particle
swarm optimiser," in Systems, Man and Cybernetics, 2002 IEEE
International Conference on, 2002, p. 6 pp. vol.3.
[30] M. Assadian, M. M. Farsangi, and H. Nezamabadi-pour, "GCPSO in
cooperation with graph theory to distribution network reconfiguration
for energy saving," Energy Conversion and Management, vol. 51, pp.
418-427, 3// 2010.
[31] S. Yang, H. Cheng, and F. Wang, "Genetic Algorithms With Immigrants
and Memory Schemes for Dynamic Shortest Path Routing Problems in
Mobile Ad Hoc Networks," IEEE Transactions on Systems, Man, and
Cybernetics, Part C (Applications and Reviews), vol. 40, pp. 52-63,
2010.
[32] D. Zhang, Z. Fu, and L. Zhang, "An improved TS algorithm for loss-
minimum reconfiguration in large-scale distribution systems," Electric
Power Systems Research, vol. 77, pp. 685-694, 4// 2007.
[33] T. T. Nguyen, S. Yang, and J. Branke, "Evolutionary dynamic
optimization: A survey of the state of the art," Swarm and Evolutionary
Computation, vol. 6, pp. 1-24, 10// 2012.

Kumar Utkarsh (S16) received the Bachelor of


Technology degree in Electrical Engineering from
the Indian Institute of Technology (BHU), Varanasi,
India, in 2011. He is currently working towards the
Ph.D. degree in Electrical and Computer
Engineering at the National University of Singapore,
Singapore. His research interests are in power
system optimization, renewable energy integration,
and utilization and development of computational
intelligence and distributed control techniques for
efficient smart grid operation.

Anupam Trivedi (M15) received the Dual degree


(Integrated Bachelors and Masters) in Civil
Engineering from the Indian Institute of Technology
(IIT) Bombay, India, in 2009. He received the Ph.D.
degree in Electrical and Computer Engineering from
National University of Singapore, Singapore in 2015,
where he is currently a Research Fellow. His research
interest is in evolutionary algorithms, multi-objective
optimization, and power system scheduling.

Dipti Srinivasan (M89SM02) received the Ph.D.


degree in Engineering from the National University
of Singapore in 1994. She worked as a Postdoctoral
Researcher at the University of California, Berkeley,
from 1994 to 1995 before joining the National
University of Singapore, where she is currently an
Associate Professor in the Department of Electrical
and Computer Engineering. Her research interest is
in the application of soft computing techniques to
engineering optimization and control problems.

2471-285X (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Potrebbero piacerti anche