Electric Power Systems Research 152 (2017) 271–283

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

A distributed minimum losses optimal power flow for islanded

microgrids
Eleonora Riva Sanseverino a , Luca Buono c , Maria Luisa Di Silvestre a , Gaetano Zizzo a,∗ ,
Mariano Giuseppe Ippolito a , Salvatore Favuzza a , Tran Thi Tu Quynh a ,
Nguyen Quang Ninh b
a
DEIM – University of Palermo, viale delle Scienze, I-90128, Palermo, Italy
b
IES – Institute of Energy Science, 18 Hoang Quoc Viet, Caugiay, Hanoi, Vietnam
c
Terna S.p.A., viale E. Galbani n.70, I-00156, Rome, Italy

a r t i c l e i n f o a b s t r a c t

Article history: In this work, the minimum losses optimal power dispatch problem for islanded microgrids with dis-
Received 18 July 2016 tributed energy resources (DER) is solved by means of a distributed heuristic approach. Optimal power
Received in revised form 3 June 2017 management is performed almost in real time, with a predefined schedule, i.e. every 5 min, and the solu-
Accepted 11 July 2017
tion is applied to generators when the current operating solution violates voltage or current constraints
or when the current configuration produces too large power losses.
Keywords:
The operating point of both inverter-interfaced generation units as well as rotating production sys-
Distributed optimization
tems can be modified simply using local information. The latter are voltage measurements and power
Microgrids
OPF
injections or loads data of local and nearby nodes, therefore information processed at each bus derive
Islanded operation from communications between adjacent nodes.
The distributed algorithm is iterative but also fast and easy to understand, since it is based on the use
of power flow equations. It can be employed for small and medium size networks showing tens of nodes
and test results prove that convergence happens in few iterations.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction The term DER comprises a variety of DG units such as pho-

Distributed generation (DG) and energy efficiency are the basis
of the innovative models of power distribution organized in energy
districts by particular smart grids (SGs), also known as microgrids
(MGs). According to the Department of Energy (United States), MGs
are a set of interconnected loads and distributed energy resources
(DERs) with clearly defined electrical boundaries that acts as a sin-
gle entity with respect to the grid and can connect and disconnect
from the grid [1]. Thanks to the development of new advances in
Information and Communication Technologies (ICT) for power sys-
tems, the intelligent integration of the actions of all the connected
units (such as generators, consumers and prosumers) is carried out,
in order to efficiently deliver economical and secure sustainable
supplies of electricity (as the smart grid concept outlined by the
‘European smart grid technology platform’ in 2006) [2].
∗ Corresponding author. Fax: +39 091488452.
E-mail address: gaetano.zizzo@unipa.it (G. Zizzo).
tovoltaic (PV), wind turbines (WTs), and energy storage systems
(ESS). Sound management of a MG requires an energy manage-
ment system (EMS) which controls at the highest level the energy
flows, inside the MG and between the latter and the main grid, by
adjusting both the energy flowing in or out from the main grid, the
dispatchable DER, and the controllable loads based on signals or
forecasts on weather and energy prices.
Such EMS, when the MG operates in islanded conditions will no
more account for the energy market but for energy production costs
and pollution. When the MG is islanded, the optimal power flow
(OPF) issue is also quite relevant, since power flows and operational
limits must be accounted for, especially if the grid was not originally
designed to host a large penetration of RESs or if the size of the
system itself is large enough to consider losses a relevant quantity
in optimizing the energy management.
Recently, OPF algorithms for smart transmission system, where
RESs are integrated, have been proposed, such as the one described
in Ref. [3]. However few papers consider the OPF problem for radial
distribution systems [4], also involving distributed generators and
storage units [5]. Even fewer papers study the issue in MGs [6–9].

http://dx.doi.org/10.1016/j.epsr.2017.07.014
0378-7796/© 2017 Elsevier B.V. All rights reserved.
272 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283
Due to the high computational burden, the centralized OPF
problem’ solution can become prohibitive in case of medium and
large size systems and not so reliable in smaller systems. For
real-time networks management, it is in general needed to find
a new network operation layout in a few seconds or minutes, so
as to promptly re-organize operation following fast load varia-
tions and intermittent power generation changes that is typical
of renewable-based DG units. Moreover, centralized control archi-
tectures are less fault tolerant than decentralized ones. The latter
can easily host new nodes and new functionalities without large
interventions enabling plug-and-play characteristics.
Distributed algorithms are proposed for transmission systems
minimizing production costs [10,11]. The work in Ref. [10], in
particular, shows an efficient distributed OPF method suitable
for self-optimizing parts of the transmission power system. In
each part, a centralized Security Constrained Optimal Power Flow
(SCOPF) is solved with constraints over tie-lines connecting to other
parts of the power system and the method, iteratively, reaches a
convergence. The method offers the novelty of using both power
quantities and boundary prices. In Ref. [11], a general framework
for distributed economic OPF in transmission power systems is
provided, but still employing complex QP solvers that are not
easily transferable to cheap distributed microprocessors. Other
optimal management distributed algorithms for MGs are devoted
to distributed resources management in a market context. In Ref.
[12], a distributed algorithm based on the classical symmetrical
assignment problem is proposed. In this case, the algorithm solves
optimally the negotiation between agents in the MG. No compu-
tation is made to minimize power losses during operation. Other
distributed algorithms are proposed for energy management in
the literature for MGs, in most cases devoted to manage energy
resources in the market context. A distributed scheduling strat-
egy for a MG system to minimize the cost of the non-green energy
consumption is considered in Ref. [13], where the scheduling prob-
lem is formulated as a privacy-constrained linear programming
problem. Also in this case, the main aim is the energy market man-
agement. In Ref. [14], several distributed algorithms are proposed.
However, they are aimed at solving the distributed resource allo-
cation (also loads) problem, facing the decentralized coordination
and control of demand response (DR) resources and DERs. Lever-
aging the dual decomposition, a distributed energy management
approach is developed in Ref. [15] for MGs with high penetration of
RESs. In this case, the framework is again a liberalized energy mar-
ket with an aggregator sending power curtailment/increase signals
to the distributed energy resources. In Ref. [16], Additive Increase
Multiplicative Decrease (AIMD) algorithms are adopted to share the
power generation task in an optimized fashion based on minimum
production cost among the DERs within a MG. No realistic technical
constraints are considered (i.e. power losses are neglected).
Several distributed algorithms can be found in the literature for
efficient OPF solution in power systems and in MGs. The PhD dis-
sertation from Cai [17] proposes different distributed multi-agent
systems for solving the OPF in microgrids. The simplest does not
consider network losses and voltage regulation. In this case, the
information flows in parallel and results are obtained in a non-
iterative way; therefore, the method is fast and does not show any
convergence issues. The same work [17] proposes a multi-agent
control for power balance or attaining economic dispatch based
on Gauss method for systems with considerable power losses and
where voltage regulation is expected. The proposed power flow
algorithm makes use of communication time, and updates state
information synchronously among agents. However, the conver-
gence is reached after hundreds of iterations and it seems that
power losses minimization is not considered within the OPF solu-
tion, moreover it cannot be applied directly to islanded microgrids
(slack bus showing limited capacity). Finally, the paper in Ref. [18]
solves the economic OPF for microgrids in a distributed manner
accounting for power balance equations as constraints. In this case,
an event-triggered distributed algorithm provides the solution. The
latter can be implemented by dynamically adjusting the power
set-point of each generator. In this way, the OPF problem can be
converted into a sequence of unconstrained problems by adding to
the cost function a penalty term that gives a high cost to unfea-
sible points. Each generator solves an optimization problem and
requests information to neighboring states only if necessary (when
event triggered). However, the approach does not account for losses
and the example shows very limited size.
In Ref. [19], a distributed minimum cost energy management
for scheduling of energy resources in a MG is proposed. In the
formulation, the optimal operation of MGs takes into account the
distribution network infrastructure and associated constraints. The
energy management problem is formulated as a minimum produc-
tion cost issue, where the MG Central Controller (MGCC) and the
local controllers jointly compute an optimal schedule. In this case,
the energy management problem is thus formulated as the solution
of an OPF problem in different times.
Recent works consider distributed computations aimed at
optimal reactive power flow management in distribution net-
works based on dual decomposition [20] and Alternating Direction
Method of Multipliers (ADMM) methods [21]. However, there are
cases for which the ADMM-based solution of the non-relaxed OPF
problem fails to converge [22,23].
The latter works analyze the overall picture of OPF underlining
the limits of the two main cathegories of OPF solution methods:
those approximating the physical network model and those relax-
ing the space of solutions and/or control variables. The same works
[22,23] recognize also a recently arisen third cathegory called
Branch Flow Model methods (BFM) for which the network flows
are described by using as variables the currents and the powers of
the various network branches, instead of the nodal injections. The
work in [23] proposes a distributed OPF that provides exact solution
and can only be used for radial systems.
Very few works in MGs literature explicitly analyze the min-
imum losses issue considering a possible islanded operation and
tuning both real and reactive power from generators. The work in
[24] offers a nice distributed OPF for grid connected MGs consid-
ering smaller partitions, but still remains the problem of defining
what is a good partition and moreover also [24] employs commer-
cial solvers hardly usable on industrial microcontrollers.
In Refs. [25] and [26], the entire centralized OPF is formulated
as a combination of sub-problems that is solved in a distributed
manner by each bus. In particular, [26] proposes a consensus-based
distributed OPF algorithm for optimal economic dispatch, which
uses consensus algorithm to estimate the optimization variables
between neighboring nodes in a given network. Agents at the buses
run the algorithm. They own a portion of the power grid repre-
sentation and solve a local optimization. The agent not only uses
the voltage and power variables measured at local level, but also
estimates the voltage and power variables coupled with the local
power balance constraints. The agents exchange their local data and
estimates with neighboring agents to create a correct estimation.
To improve the estimation, a consensus filter is used to process the
data shared among nearby agents. A penalty function-like method
is used to minimize the local objective. However, the number of
iterations seems to be quite large (200 iterations for 14 bus sys-
tem), thus affecting computation time. Moreover, it is not clear
when output power from generation buses/storage units should be
changed, since there is no prove that all buses reach convergence
at the same time. In Ref. [25], by means of an ADMM-based dis-
tributed algorithm using closed form solutions to the optimization
sub-problems, fast and distributed solutions for real-time feed-
back control are obtained for balanced radial networks. However
 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283
Table 1
Comparison of minimum loss DOPF algorithm for MGs.
Approach Grid con- Commercial Generators
nected solvers Synchronization
Reference [23] Y N Dealt with
Reference [24] Y Y Not explained
Reference [25] Y/N N Not explained
Reference [26] Y Y Not explained
Proposed DOPF Y/N N Dealt with
the number of iterations still remains high and, such as in most
papers dealing with distributed approaches, it is not clear how
synchronization issues between nodes are handled as well as how
(under what triggering event) and when generators outputs can be
modified. Table 1 shows a comparison of some relevant features
of some distributed OPF algorithms in the literature. They are all
comparable since dealing with microgrids and solving OPF in a dis-
tributed fashion either aiming at minimization of losses or aiming
at the minimization of production costs. Some of them only pro-
pose grid connected examples and do not highlight the possibility
that the algorithm can be used for islanded systems. Even in this
case, most of them use commercial solvers thus not allowing the
easy implementation of the methodology over microcontrollers. In
most papers, it is not clear when and how the output powers from
generators should be applied (generators synchronization). More-
over, not in all cases the level of synchronization reaches the node
level.
In the present paper, the problem of optimal management
of a smart power distribution grid using a distributed heuristic
approach is proposed. The OPF method here proposed uses a net-
work description based on the Branch Flow Model [23].
In this paper, which is the full development of a study started
by the same authors in recent times [27], there are four main con-
tributions.
Firstly, it proposes a distributed intelligence algorithm for small-
medium non-radial distribution systems OPF. It has been conceived
for islanded systems, but can easily be employed for grid-connected
networks.
Secondarily, it proposes a plug-and-play system for connecting
micro-sources and loads to the grid, without any need of recon-
figuring the Micro Grid Central Controller (MGCC). The system is
fault-tolerant and can still work if one or more units experience
both a software and a hardware failure, provided there is commu-
nication between nearby nodes.
Thirdly, the algorithm in a few iterations learns a set of weights
that account for constraints and for the relevance of some power
flows in certain branches.
Finally, differently from all other approaches the algorithm pro-
poses a way to synchronize the generators output variations in
order not to determine unbalances and associated instabilities.
As compared to a large part of existing literature on the topic,
the approach proposed by the authors in this paper delivers a solu-
tion for the new set-points of the generation units in a short time
and with an easy approach. The topology of the microgrid can be
meshed, while most of the approaches listed in Table 1 and appear-
ing in the literature only account for radial architectures. As in
Refs. [23,25,26], also in this work communication is only required
between adjacent buses. As far as practical implementation is con-
cerned, in the algorithm, a way to timely organize the new output
power set-points application is deployed. Only when these receive
a final feedback from the sink buses that are considered those where
273
CPU time order of Level of decentralization Objectives
magnitude
Number of exchanged messages Node Minimum losses/economic
between nodes: thousands dispatch
Few seconds/Intel Core i7-2600 CPU @ Area (sub-networks) Minimum losses/Economic
3.40GHz dispatch
Second and below/not specified Node Minimum losses/Economic
dispatch
Not specified Node Economic dispatch
Seconds/Intel Core i7-2600 CPU @ Node Minimum losses
3.40 GHz
Number of exchanged messages: tens
the verification about convergence is carried out, they will apply the
new operational set-point. Moreover, in order to avoid a loss of syn-
chronization during the execution of the algorithm, it is imposed
that each bus starts calculating, based on its own function, only
when it receives data from all adjacent buses.
The authors have focused their attention on power losses mini-
mization because this objective seems the toughest to be achieved.
On the other hand, in MGs, where droop controlled inverter-
interfaced units are present, the possibility to move from one
minimum loss operating point to another following closely loads
change will improve the overall system performance.
A main source of error in the considered DOPF is the represen-
tation of the meshed system as a set of radial systems divided by
only active power inversion nodes. While the convergence of the
B/F method, on which the methodology is based for calculations,
although guaranteed in a lot of cases as proved in Ref. [35] can
experience a slow down due to the magnitude of the equivalent line
impedance and load admittance of the branches of the network.
The rest of the paper is arranged as follows. Section 2 describes in
details the formulation of the distributed OPF problem and the solu-
tion algorithm is reported in Section 3. Section 4 shows applications
of the proposed algorithm on smaller and larger power distribution
systems. Finally, results are commented and future directions are
pointed out to the reader in Section 5.
2. Problem formulation
The optimal power dispatch among DERs, or OPF issue is a non-
linear constrained problem in which, typically, variables are the set
points of generation systems and, in many cases, the optimization
is carried out at the MGCC in a centralized fashion. Problems with
most of the existing distributed approaches [12–16,19], arise due
to two reasons: (1) communication between MGCC and peripheral
units (load controllers, LC, and DG or ESS) is needed and (2) these
approaches consider the supply-demand matching in an abstract
way, where the aggregate demand is simply equal to the supply.
As far as the first issue is concerned, the need for a hierarchical
architecture for communication and control shows all the draw-
backs of centralized architectures and do not easily allow future
expansions or modifications of the grid.
As far as the second issue is concerned, in many cases, all the
generation and load units are modeled as if they are connected to a
single grid bus, neglecting the underlying physical behavior of the
power distribution network, namely associated power flows (e.g.:
Kirchhoff’s law) and system operational constraints (e.g.: voltage
and branches ampacity tolerances). As a result, the schedules pro-
duced by those algorithms may violate those constraints and thus
turn to be not feasible in practice.
The main hypotheses underlying the approach proposed in this
paper are the following:
274 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283

1) each bus can have information both locally and from adjacent
buses and branches;
2) only voltage modules and bus power supply/demand can be
measured.

The branch information is resistance, reactance and ampacity

of the line. Each bus processes the data and then sends results to
Fig. 1. Generic bus at which the DOPF is solved.
adjacent buses for further elaborations only when it receives data
from all adjacent buses.
Losses are determined solving a centralized load flow as it fol-
lows:
2.1. Hierarchical MG control architecture and classical

G

L

B
centralized OPF PiS = PiL + Pk (2)
i=1 i=1 k=1
With the basic hierarchical control architecture proposed in the
literature for MGs [38], when a bus power injection suddenly varies
(from a load or a generation bus), the regulation process starts. In

G 
L 
B
QiS = QiL + Qk
MGs control structure [38], three different control levels can be
i=1 i=1 k=1
evidenced:
under the following constraints:
– Primary control level is typically used for controlling local power, Vimin ≤ Vi ≤ Vimax i = 1, . . .N (3)
voltage and current acting on power conversion systems. This
control level typically follows the set points given by upper level Ik ≤ Ikmax k = 1, . . .B (4)
controllers;
PiS,min ≤ PiS ≤ PiS,max i = 1, . . .G (5)
– Secondary control level manages power quality control, such as
voltage/frequency restoration, as well as voltage unbalance and where:
harmonic compensation. In addition, it takes care of synchroniza-
tion and power exchange with the main grid or other MGs; • Vi is the voltage module at the sending bus i of branch k; i = SB(k);
– Tertiary control level finally attempts to optimize the MG • Pkflow , Qkflow are the real and reactive power flows on branch k,
operation based on efficiency and/or economics, solving when respectively;
necessary the OPF problem. Knowledge both from the MG • Ik , Imax
k are the current flow module on the k-th branch and the
side as well as from the external network is of great impor- k-th branch ampacity, respectively;
tance to optimize the different functions. Local or centralized • Vi , Vmin
i , Vmax
i are the voltage module and its minimum and max-
Decision-Making algorithms are employed to process the gath- imum rated values at bus i, respectively;
ered information and take proper actions. • PSi , PS,min , PS,max are the real power injection at the i-th dis-
i i
patchable generation node, its minimum and maximum values,
respectively.
The communication channels bandwidth of the various control
levels are typically different by at least one order of magnitude. In
If the controllable generator unit is a storage unit, the power
this way, the dynamics at different levels can be considered decou-
limits in Eq. (5) refer to the maximum charge or discharge powers
pled and this simplifies to a great extent modeling and analysis.
calculated considering the available State of Charge.
At higher control levels, regulation speed is lower; e.g. droop
The optimization variables are the real and reactive power at the
control in primary level has typically a response time between
generation buses, therefore Eq. (1) must relate to these quantities.
1 ms and 10 ms, while secondary control speed can stay between
100 ms and 1 s, based on the speed limit of the used communi-
2.2. The decentralized formulation for OPF problems
cation technology. Finally, tertiary control implements the actions
in time steps between seconds and hours. While for lower control
This section provides a decentralized formulation of the same
levels, extensive literature provides decentralized implementation,
problem. Consider a generic bus BE as represented in Fig. 1.
decentralized approaches for tertiary regulation level are still under
The function to be minimized locally is expressed through the
investigation. At tertiary level, the solution of the OPF provides
function f defined below, that expresses the summation of active
new operating set points for generators giving rise to minimum
power losses of the branches connected to a generic ending bus,
losses operation and, with limited effort, also to minimum cost
showing at least two branches with power flows entering the same
operation. This paper proposes the identification of a sub-optimal
node:
minimum losses operating point for generators using a distributed
intelligence methodology.
 RBE−BS(i)
Nji (BE)
2 2
f = 2
. · (ωpi · PBE−BS(i) + ωqi · QBE−BS(i) ) (6)
The standard OPF issue for power loss minimization can be VBE
i=1
expressed as follows. Consider a N bus MG, where G out of the N
buses are generators, including storage systems, and L out of the N with Nji (BE) ≥ 2where:
buses are loads or non-dispatchable RESs. The MG has B branches,
for each of which the longitudinal electrical parameters can be indi- • BE: ending bus;
cated as Rk and Xk (where k = 1,..,B). The mathematical formulation • BS(i): i-th sending bus connected to node B
of the centralized OPF problem can be written as it follows: • RBE−BS(i) : resistance of the branch connecting BE and BS(i);
• VBE : voltage module at bus BE;

B 
B
R 2 2
• Nji (BE): number of sending nodes adjacent to bus BE;
k k k
Min Pk = Min 2
[(Pflow ) + (Qflow ) ] (1) • PBE−BS(i) , QBE−BS(i) : real and reactive power flows afferent to bus
V
k=1 k=1 i BE (local optimization problem variables).
 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283 275

• ωpi : normalized weight associated to the active power flow in the

i-th branch, connected to the bus BE;
• ωqi : normalized weight associated to the reactive power flow in
the i-th branch, connected to the bus BE;

and with the following constraints:


Nji (BE) 
Nij (BE)
PBE−BS(i) = (Pi BE ) + PLBE (7)
i=1 i=1

Nji (BE)

Nij (BE)
QBE−BS(i) = (Qi BE ) + QLBE (8)
i=1 i=1

As shown in Eq. (6), two weights have been associated to the

Fig. 2. Polling process during the tracing procedure.
branch power flows also accounting for the upstream losses. With-
out these, the objective function would only express local losses and
thus the local optimum achieved by minimizing Eq. (6) considering
one in the j-th column if there is an electrical connection between
the constraints Eqs. (7) and (8) could not bring a reduction of over-
bus i and bus j, otherwise it shows a zero. The IED at bus i only
all power losses. This weighted sum formulation allows generating
has access to the i-th row, meaning that it has a ‘local’ view of the
solutions that account for the upstream losses. Therefore, looking
topology of the network.
at bus BE, if one of the upstream branches shows a high resistance,
The p-th bus stops the polling process when the following con-
the power flow can take a higher value, thanks to the introduction
dition holds:
of weights that account for the global optimization. As far as con-
straints are concerned (see conditions Eqs. (3)–(5) in centralized
Nji (p) + Nij (p) = N (p) (9)
OPF formulation), they have not been mathematically considered,
since they would put a computational overhead to local solvers.
where Nji (p) is the p-th entry of the vector Nij containing the num-
These constraints have been accounted for ‘heuristically’ acting on
ber of branches showing a power flow entering bus (p), while Nji (p)
weights and using some feedbacks, as it will be detailed below.
contains the number of branches that are electrically connected to
the p-th bus with power flowing out of bus p. N(p) is the number
3. Heuristic distributed algorithm
of buses electrically connected to bus p. In Fig. 2, an example of the
polling process is outlined for a 9 buses system.
In the following sections, the iterative algorithm is detailed.
Once the polling process is terminated, the calculation of the
For its solution, the visit of all branches of the network is needed,
features of interest and the feedbacks propagation is executed. Each
because each bus has a partial knowledge of the network. For this
generator is indeed characterized by quantities named ‘feedbacks’
reason, the solution is reached in a limited time, according to the
that will be defined in the following sections.
time required by the available telecommunication infrastructure
A bus that is enabled to perform calculations is a bus at which
to visit the entire network a few times. The proposed algorithm is
a given ‘enabling condition’ is true. The p-th bus is enabled if
divided into three procedures, described extensively below:
the counter C(p) takes the value of Nji (p). Every time the p-th
bus receives information from an adjacent node, the counter C(p)
– Tracing procedure;
increases of one. Therefore the buses not showing adjacent sending
– Backward procedure;
buses, such as generation buses, are those that initiate the calcula-
– Forward procedure.
tions. For them indeed the following condition holds:

The Tracing procedure is executed only once while the back-

ward and forward procedures are executed iteratively till when
the stopping conditions are verified.
3.1. Tracing procedure
The aim of the Tracing procedure is the initial quantification
of the real and reactive power flows and their directions in the
network branches, based on voltage modules measurements. The
procedure also aims at the initial evaluation of the generators
‘feedbacks’, that can be used to efficiently handle constraints on
generated power.
The procedure starts at the generator nodes, which ask the adja-
cent nodes about their voltage level. Such polling is carried out for
all branches of the network. The result of this procedure is the iden-
tification of a Sending Bus (BS) and an Ending Bus (BE) for each
branch. In order to perform the procedure, each Intelligent Elec-
tronic Device (IED) at each bus needs to have a local knowledge
about the network topology. From a practical point of view, this
problem is solved by setting the input and output ports of each IED
as well as communication protocol details. To simulate this local
knowledge at software level, an adjacency matrix (T) showing n bus
rows and n bus columns is built. The i-th row of the matrix shows a
C(p) = Nji (p) = 0 (10)
The propagation of the information follows the real power
flows direction. The information are generated starting from the
generator buses, without adjacent sending buses, and stop their
propagation at the sink buses, namely buses without adjacent end-
ing buses.
3.1.1. Initial power flows and losses calculation
The buses that are enabled must execute the power flows and
losses calculation in the branches with power flowing out of the
bus. It is well-known that the status of the power system can be
determined starting from the modules and displacements of the
bus voltages. Nonetheless, one of the underlying hypotheses is that
each IED only has a view over the bus voltage modules. Actually,
the problem of power flows and losses calculation is not difficult
if the enabled bus is a sending bus only showing one ending bus,
since the power injected at the bus BE is the branch flow, as shown
in Fig. 3.
In what follows, the expressions for the calculation of the fea-
tures of interest for the power flow and losses calculation in the
branch BS-BE are shown. Another underlying hypothesis (valid at
276 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283


Nji(BS)

Q = (QkBS ) + QGBS − QLBS (19)

k=1

Therefore:

PBS−BE(k) = ˛k · P; QBS−BE(k) = ˛k · Q (20)

The sum of the coefficients ␣k for the branches that are afferent to
bus BS with flow coming out of BS must equal 1, in agreement with
the following condition:
Fig. 3. Generator bus with one adjacent BE. 
Nij(BS)

Nij(BS)

Nij(BS)

P= PBS−BE(k) = (˛k ) · P → (˛k ) = 1 (21)

k=1 k=1 k=1

Nonetheless, following the approximations executed, it turns

that the sum of the coefficients is slightly below the unity. Therefore
according to the calculation (Eq. (17)) the following normalization
can be carried out:
˛k
˛k = Nij(BS) (22)
(˛i )
i=1
Once the real and reactive power flows entering the k-th branch
are known, it is possible to calculate the power losses in the branch
Fig. 4. Generator bus with more adjacent BE.
as well as the active and reactive power flows afferent to bus BEk
through Eqs. (13)–(16).
distribution level) is that the branches show concentrated series
Then the IED at bus BS, once enabled, must calculate the real
parameters and null shunt parameters:
losses seen from the bus BEk , using the information propagated to

Nji(BS) bus BS from the sending adjacent buses. As an example, taking the
PBS−BE = PBS
G + (PBS BS
k ) − PL (11) situation in Fig. 2, if bus 4 is supplied by generator 1, it will compute
k=1
the power losses from the summation of the losses in branch 1-7
and those in branch 7-4.

Nji(BS)
The knowledge of losses seen from the sink buses up to the gen-
QBS−BE = QBS
G + (QBS BS
k ) − QL (12) erators is required to calculate two normalized weights, ωpi and
k=1 ωqi , that, as outlined in Section 2, are employed for the solution of
the constrained problem, as it will be detailed in subsequent sec-
(PBS−BE )2 + (QBS−BE )2
PBS−BE = RBS−BE · (13) tions. The use of the normalized coefficients ␣ implies an error in
V2BS the evaluation of the flows and of power losses in the branches
supplied by the considered bus. Nonetheless, the aim of the tracing
(PBS−BE )2 + (QBS−BE )2
QBS−BE = XBS−BE · (14) algorithm is also to estimate the power losses for the calculation of
V2BS ωpi and ωqi .
PBE−BS = PBS−BE − PBS−BE (15)
BE−BS BS−BE
3.1.2. Generators feedback propagation
Q =Q − QBS−BE (16) In most cases, MGs host renewable and non-dispatchable power
In these expressions, PG BS and QG BS are the generated real and reac- generation units. For these units, the supplied power cannot be con-
tive powers injected at bus BS, while PBS-BE and QBS-BE are the trolled. These sources may become dispatchable only if suitably
real and reactive losses at the branch BS-BE, respectively. If the connected to a storage unit [29]. In order to account for these not
enabled bus shows more than one adjacent ending bus (BE1 , BEk , controllable power injections, as well as for a power flow between
BENij(BS) ), the problem of the partition of the entering flows into the MG and the main grid which is contractually constrained, a spe-
more outgoing flows, such as (BS-BEk in Fig. 4), arises. cial procedure must be adopted. To satisfy this equality constraint,
The problem is intrinsically undetermined, since the partition during the tracing procedure, the non-dispatchable generators
of the flows depends on the modules and displacements of the have assigned a feedback that propagates to adjacent nodes. In this
bus voltages. Nonetheless, in distribution networks the voltage dis- way, the constrained flows are identified and will not be modified
placements are all very close and the differences can be neglected. during the optimization. Once the feedback is generated, it must
Based on this simplifying assumption [28,39], the power can be propagate till when it meets a sink bus or a bus with at least two
distributed among the branches afferent to node BS in which the adjacent sending buses, with at least one unconstrained bus. The
power flow is spreading out of the bus (see Fig. 4), according to feedback may take one of these two values:
real coefficients ␣, that are associated to each of these branches.
Referring to the branch BS-BEk in Fig. 4, it can be written: 1) +1, if the constrained bus propagating the feedback has only one
adjacent ending bus, Nij (BS) = 1;
VBS 2 − VBEk · VBS 2) −1, if the constrained bus propagating the feedback has more
˛k = (17)
(RBS−BEk · P + XBS−BEk · Q ) than one adjacent ending bus, Nij (BS) > 1.
with k = 1,. . .,Nij (BS) and Nij (BS) ≥ 2;
where: A feedback equal to +1 means that the power flow in the branch is
constrained and known without approximation. A feedback equal

Nji(BS)
to −1 means a partially constrained power flow. As a result, the
P= (PkBS ) + PGBS − PLBS (18) constrained branches are not subjected to optimization, while the
k=1 partially constrained power flows are subjected to optimization,
 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283 277

only at the first iteration of the Backward procedure. Such approach Executing the needed substitutions in the system’s Eq. (25), a linear
is coherent with the power losses minimization issue: when a system can be attained as it can be represented under the following
constrained generator supplies a bus with at least two afferent matrix form:
⎡ ⎤
branches, it is possible to identify an optimal operating condition PBE−BS(1)⎡ ⎤
fulfilling the equality constraint about the power injected by the 0
⎢ ⎥
generator, minimizing the losses in the branches supplied by the ⎢ PBE−BS(i) ⎥ ⎢ . ⎥
⎢ ⎥ ⎢ .. ⎥
same generator. ⎢P ⎥ ⎢ ⎥
⎢ BE−BS(n) ⎥ ⎢ 0 ⎥
As it will be detailed in the following, starting from the second ⎢ ⎥ ⎢ ⎥
iteration of the Backward procedure, the branches that were at first ⎢ QBE−BS(1) ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
only partially constrained become fully constrained, so as to meet R· ⎢ ⎥=⎢ ⎥ (26)
⎢ QBE−BS(i) ⎥ ⎢ .. ⎥
the equality constraint about generated power. Referring to Fig. 5, ⎢ ⎥ ⎢.⎥
⎢ ⎥ ⎢ ⎥
an example of constrained generation for DG1 (DG Unit #1) and ⎢ QBE−BS(n) ⎥ ⎢ 0 ⎥
relevant feedbacks propagation is given, when bus 7 is supplied by ⎢ ⎥ ⎢ ⎥
⎢  ⎥ ⎣P⎦
the three generators. ⎣ 1 ⎦
Q
2
with:
3.2. Backward procedure
• BE: enabled bus;
The Backward procedure aims at the determination of the power • BS(i): i-th sending bus adjacent to BE;
flowing in the network branches for which a real power losses • PBE−BS(i) , QBE−BS(i) : real and reactive flows of the branch BE-BS(i)
reduction is attained, following the bus Kirchhoff law. For a cor- connected to BE with flow entering the bus (optimization vari-
rect synchronization of the algorithm, also the Backward phase is ables);
characterized by the presence of counter C(p), that enables the exe- • P,Q: equivalent real and reactive flows downstream the bus BE,
cution of calculations at the p-th bus if and only if it takes value respectively obtained as the summation of active and reactive
equal to Nij (p). The enabling variable is subjected to update every flows spreading out of the considered bus, minus the reactive
time the p-th bus gets information from an adjacent ending bus. power required by the load at the same bus.
Therefore, the buses that are automatically enabled are sink buses, • n = Nji (BE);
for which the following relation is satisfied: • R: matrix of coefficients of order 2n + 2.

C (p) = Nij (p) = 0 (23) where:

 
W ATeq
R= (27)
Therefore, differently from the tracing procedure, the informa- Aeq Z
tion propagation is in the opposite direction compared to the real
power flows. In particular, the Backward procedure originates at ⎡ ⎤
A1 · ωp1 0 ... ... ... 0
the sink buses and terminates at the generator buses without adja-
⎢ ⎥
cent sending buses. The Joule losses minimization is carried out by ⎢ ⎥
⎢ 0 Ai · ωpi 0 ... ... 0 ⎥
means of the nodes for which Nji (p) > 1. These behave as real and ⎢ ⎥
reactive power valves and they modulate the downstream power ⎢ An · ωpn ⎥
W =⎢ ⎥
0 0 0 ... 0
flows. ⎢ ⎥
⎢ 0 ... ... A1 · ωq1 0 0 ⎥
In what follows, the calculations to be executed by the enabled ⎢ ⎥
buses BE during the Backward procedure are outlined. ⎢ ⎥
⎣ 0 ... ... 0 Ai · ωqi 0 ⎦
0 ... ... ... 0 An · ωqn
Ri
3.2.1. Optimization problem solution withAi = 2 · 2
VBE
The solution of the optimization problem (Eq. (6)) only with (28)
equality constraints, under the hypothesis that the objective func-
tion f can be differentiated, can be executed by means of the
 
Lagrange multipliers [30–32]. These turn the constrained problem 1 ... 1 0 ... 0
into an unconstrained one. Aeq = (29)
0 ... 0 1 ... 1
Once the Lagrangian function has been introduced:  
0 0
Z= (30)

p 0 0
L(x, ) = f (x) + i · ci (x) (24) W is a 2n × 2n matrix, Aeq is a 2 × 2n matrix and Z is a 2 × 2 matrix.
i=1 It can be easily proved that the solutions of the optimiza-
tion problem are inversely proportional to the associated weights.
it is required to solve the following system of n+p equations under Therefore, the greater the weight, the smaller the power and vice
the n unknowns xi and the p unknowns i : versa. This issue must be considered for the weights update. It can
be observed that the order of the system (26) is tightly connected
⎧ to the number of sending buses adjacent to the bus BE.
⎪
⎨ 
p
The minimum size of the linear system can be attained for
∇ x L(x, ) = ∇ x f (x) + i · ∇ x ci (x) = 0 Nji (BE) = 2. In the truth, not all branches with flow entering the bus
(25)
⎪
⎩ i=1 are part of the optimization problem. As already mentioned in Sec-
∇  L(x, ) = c(x) = 0 tion 3.1.2, the constrained branches showing a feedback equal to 1
278 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283

Fig. 5. Constrained generation at bus DG1 and relevant feedbacks propagation.

are not subjected to optimization and therefore the flows afferent If, for the branch BE-BS(i), the real power flow inversion does
to bus BE from these branches must be considered known terms to not take place, the following relations can be used:
be brought in the second member of Eqs. (7) and (8), and only the 2
(PBE−BS )2 + (QBE−BS )
remaining unconstrained branches are subjected to optimization. PBS−BE = RBS−BE · 2
(31)
With reference to this situation, some cases can be distinguished. VBE
2
(PBE−BS )2 + (QBE−BS )
1. If only one of the branches with entering flow is not constrained, QBS−BE = XBS−BE · 2
(32)
VBE
it is enough to apply the Kirchhoff law for the determination of
the power flowing in it. PBS−BE = PBE−BS + PBS−BE (33)
2. When more than one branch shows a flow entering a bus, then
the optimization problem must be solved. If at the considered QBS−BE = QBE−BS + QBS−BE (34)
bus also constrained branches bring power and the total sum- If, for the branch BE-BS(i) the real power flow inversion has
mation of real and reactive flows afferent to the bus is greater taken place, once the losses were calculated using Eqs. (31) and
than the equivalent real and reactive flows downstream the bus, (32), the following equations can be used:
respectively P and Q in Eq. (25), the real power flow of one of the
unconstrained branches must change its direction. In this case, PBS−BE = PBE−BS − PBS−BE (35)
the branch showing higher power losses looking upstream will QBS−BE = QBE−BS − QBS−BE (36)
be chosen for a change of direction in the power flow.
3. If no branch with entering flow is constrained, the system Eq. Analysing cases (1), (2) and (3) listed in the previous subsec-
(25) can be normally solved. tion, the real power flow inversion takes place due to a negative
active power flow, P* BE-BS(k) , or due to the inversion of the flow in
Nonetheless, if one of the branches with flow entering the bus an unconstrained branch. Following the inversion of the real power
(constrained or unconstrained) shows a feedback associated to the flow, it is required to update the matrix T and the vectors Nji and
violation of a branch current constraint violation or a maximum Nij as follows:
power from upstream generators constraint, all branches are sub- ⎧
⎪ T (BE, BS (k)) = −1 · T (BE, BS (k))
jected to optimization to allow the elimination of the constraint ⎪
⎪
⎪
⎪ T (BS (k) , BE) = −1 · T (BS (k) , BE)
violation and consequently of the feedback. ⎪
⎪
The feedbacks associated to the latter constraints violation are ⎪
⎨ Nji (BE) = Nji (BE) − 1
generated in the forward procedure, described below. (37)
⎪ Nij (BE) = Nij (BE) + 1
⎪
⎪
⎪
⎪
3.2.2. Power flows and losses calculations ⎪
⎪ Nji (BS (k)) = Nji (BS (k)) + 1
At the end of the optimization problem solution, the losses and
⎪
⎩
Nij (BS (k)) = Nij (BS (k)) − 1
power flows calculation can be carried out at buses BS(i), with
i = 1:Nji (BE). being BE-BS(k) the branch where the flow inversion took place.
 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283
The flow inversion generates new nodes in the system where
the optimization problem must be solved. At the current iteration
of the Backward procedure, the branches in which the flow has
changed its direction are not unknowns of the problem and can
therefore be brought in the second term of the equality constraints
of Eqs. (7) and (8).
The enabled buses with Nji (BE) = 1 must not solve any optimiza-
tion problem; they must calculate the power flows and losses of the
only branch showing a flow entering the node, BS-BE. In particular,
once the flow afferent to node BE of the branch has been calculated
according to the following expressions, the remaining features are
calculated by means of Eqs. (31)–(34).

Nij (BE)
PBE−BS = PiBE + PL BE
i=1

Nij (BE)
QBE−BS = QiBE + QL BE
i=1
Moreover, if the branch BE-BS is partially constrained it is
required that the bus BE verifies whether the branch downstream
is or not constrained.
If the branch is constrained, also the branch BE-BS gets con-
strained as well, while if none of the downstream branches is
constrained, also the branch BE-BS must become unconstrained.
3.2.3. Application of the optimization problem to generator nodes
When a generator with Nji (BE) > 0 is connected to the consid-
ered node, the algorithm is not applicable, since it is required to
define what amount of power downstream the bus (as a solution
of the optimization problem for the buses downstream the genera-
tor) must be assigned to the considered generator. To overcome this
problem, it is possible to use a fictitious branch, with small longitu-
dinal impedance, connecting the generator to the bus where other
downstream branches are connected.
At the first iteration of the Backward procedure, the weights
associated to the fictitious branch are imposed. The initial value of
these weights is chosen as the smallest among those of the branches
with flow entering the considered generator bus. Obviously, since
the fictitious branch has very low resistance, also the associated
losses will be small and thus, the presence of a small weight could
lead to a high value of the generated power, may surpassing the
technical threshold value. Thanks to the learning algorithm, it is
possible to handle all the weights involved in the algorithm includ-
ing those about the constraints on generators upper and lower
limits.
3.2.4. Learning algorithm
At the first iteration of the Backward procedure, the weights ωpi
and ωqi , associated to the branches connected to the ending bus
(BE), whose IED must solve the optimization problem, are equal to
the normalized value of the real power losses seen from the i-th
branch in the upstream direction, LPt(BE,BS(i)).
LPt (BE, BS (i))
ωpi = ωqi = Nji(BE)
j=1
LPt (BE, BS (j))
with i = 1,...Nji (BE) ed Nji (BE) ≥ 2
Starting from the second iteration, the weights are updated by
means of a learning algorithm. The latter modifies the weight of
the preceding iteration by an exponential factor whose exponent
is constant in module but variable in sign, the latter depending on
the feedbacks generated in the forward procedure. In particular, the
279
weights of the real and reactive flows can be different and can be
updated following the rule below [31,37]:

ωpi,IT = ωpi,IT −1 e±
(41)
ωqi,IT = ωqi,IT −1 e±
where:
• IT: iteration number, IT > 1;
• ωpi,IT , ωqi,IT : real and reactive powers weights in the current iter-
ation, respectively;
• ωpi,IT−1 , ωqi,IT−1 : real and reactive powers weights in the preced-
ing iteration, respectively;
• ␮: exponent.
(38)
The value of the exponent is chosen equal to 0.4. This value is
set by the user and, by this value, the entity of the weights increase
(39) or decrease can be set as well.
Immediately after the weights update, it is required to proceed
with the normalization, in agreement with the following expres-
sions:
⎧ ωpi,IT
⎪
⎨ ωpi,IT = Nji(BE)
i=1
ωpi,IT
ωqi,IT (42)
⎪
⎩ ωqi,IT = Nji(BE)
ωqi,IT
i=1
In this way, it is possible to increase or decrease the two power
flows in a totally independent way. Following this observation, the
weights are updated according to these criteria:
- the real power weight implements also the upper and lower con-
straints on generated power; the weight is indeed increased or
decreased according to the signs of the feedbacks generated in
the forward phase as detailed in Section 3.3.3;
- the reactive power weight implements the constraints on the
reactive power limits of generators and on voltage levels, increas-
ing or decreasing the weight according to the sign of the feedbacks
generated in the forward phase;
Finally, if there is no constraint violation, the weights will
change their value according to power losses. They will increase
their value if the feedback takes value 1 or keep the same value in
all other cases.
In general, the above cited criteria give a higher priority to the
elimination of constraints violation. Only when constraints are no
more violated, the weights modification will help power losses
reduction.
3.3. Forward procedure
Once the Backward procedure is ended, the update of the gen-
erated powers is applied and the stopping conditions are verified.
If these conditions are met, the optimization algorithm stops,
otherwise the forward procedure can start. During the forward
procedure the following quantities are updated:
– bus voltage modules;
(40) – active losses on branches;
– generated power and propagation of feedbacks for constraints;
– reference power for constrained branches.
Also the forward procedure is characterized by an enabling con-
dition that at each bus must be met for the relevant calculations.
The enabling condition is quite similar to the one described for the
tracing algorithm. In what follows, the calculations to be executed
by the BS nodes during the forward procedure are outlined.
280 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283

Fig. 6. The power flows at the generation buses.

Referring to Fig. 6, the analytical expressions for the power flows

calculations at the generation buses are reported: Fig. 7. Power flow from generation bus to sink bus.
⎧
⎪
⎪ 
Nij(BG)

Nji(BG)
⎪
⎪ = + BG
−
⎪
⎨
P G PBG−BE(i) PL PBS(j)−BG bus voltages of the BE buses of PQ type is executed in agreement
i=1 j=1 with the following expression:
(43)
⎪
⎪  Nij(BG)
 Nji(BG)
⎪
⎪ QG = QBG−BE(i) + QBG V̄BE = VBS − V̄BS−BE = VBS − ŻBS−BE ĪBS−BE (46)
⎪ L − QBS(j)−BG
⎩  
i=1 j=1
2  2
|V̄BE | = Re V̄BE + Im V̄BE (47)
3.3.1. Stopping condition
In centralized optimization algorithms, the stopping condition ĪBS−BE = |ĪBS−BE | · e−j<ĪBS−BE (48)
is often about the reduction of power losses in two subsequent
iterations. Such condition cannot be applied to the considered dis- with:
tributed algorithm. No bus in the network has knowledge about Q 
< ĪBS−BE = tg −1
BS−BE
the overall losses. Therefore a condition about the power losses (49)
PBS−BE
reduction branch by branch is in this case considered:

|LPIT − LPIT−1 | ≤ εLP (44) PBS−BE

|ĪBS−BE | = (50)
VBS · cos(< ĪBS−BE )
with:
If the bus BE is a PV bus, its voltage is not subjected to any update.
• LPIT : matrix containing the active power losses on the branches
at iteration IT;
3.3.3. Generation and propagation of the feedbacks during the
• LPIT−1 : matrix containing the active power losses on the branches
forward procedure
at iteration IT-1;
During the forward procedure, some feedbacks are generated:
• εLP : tolerance about losses decrease.

In order to prevent a premature convergence of the algorithm,
to which large errors in bus voltages estimation may correspond, a
further stopping condition about the variation of the bus voltages
at two subsequent iterations was also introduced:
|VIT − VIT −1 | ≤ εV
Where:
VIT : vector containing the bus voltages at iteration IT;
VIT−1 : vector containing the bus voltages at iteration IT-1;
εV : tolerance about voltage modules variation.
If both conditions Eqs. (44) and (45) hold, the algorithm can stop
only if no generator has surpassed its own technical limit and no
branch current has surpassed the admissible threshold.
Therefore, if the three conditions Eqs. (43)–(45) are met, the sink
bus propagates a positive feedback to the adjacent sending buses,
otherwise it sends a feedback equal to zero. The collection of all
feedbacks reaches one of the three generators, which enables the
stopping condition in all the network.
3.3.2. Bus voltage modules updates
The update is executed taking the voltage of the sending bus (BS)
of the generic branch as reference on the Gauss plane and calculat-
ing the module of the phases of the voltage at the ending bus (BE).
In this way, it is not required to know the arguments of the voltages
of the ending buses of the branch, but only their displacement.
With reference to the equivalent longitudinal circuit of the
generic branch of the network, the update of the modules of the
- bus voltage feedbacks that are sent from each sending bus to the
adjacent ending buses;
- generator real power and generator reactive power feedbacks
that start from generator buses and are propagated in the net-
work;
(45) - branch currents feedbacks that originate from any sending bus.
The feedback takes value 0, if the variable is within bounds;
value 1, if the variable overcomes the maximum admitted value;
value −1, if the value of the variable is lower than expected lower
limit. In addition, a feedback on active power seen from down-
stream nodes is required to verify whether the weights have
brought the algorithm towards losses reduction or not. It takes
value 1 if losses are reduced as compared to the preceding iteration,
−1 otherwise.
All the above listed feedbacks are used starting from the second
iteration of the Backward procedure for branches weights update.
In particular, if during the forward procedure one of the branches
upstream a given sink node is overloaded with current, the weights
update performed in the Backward phase starting from the cited
sink bus will account for this giving rise to a reduction in the
power generated from the generator bus supplying that area. Such
reduction follows indeed the reduction of the power flowing in the
branch entering the sink bus. In Fig. 7 the thicker line represents a
current ampacity constraint violation, acting on the flows in all the
lines belonging to this path, and ultimately from DG1, will reduce
the power flow also in this branch.
 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283 281

Table 2 Table 6
Electrical parameters in pu of the 9-bus test system. Vb = 10 kV, Ab = 1 MVA. Initial Operating point for the 9-bus test system.

Branch R [pu] X [pu] Ix [pu] Bus V [pu] PG[pu] QG[pu] PL[pu] QL[pu]

1-7 0.00692521 0.08702493 8.0540 1 1.1090 3.1213 0.1334 0 0

2-8 0.00692521 0.08702493 8.0540 2 1.1056 0.7097 0.4767 0 0
3-9 0.00692521 0.08702493 8.0540 3 1.1058 0.7064 0.5632 0 0
7-4 0.01288089 0.00084849 1.611 Real power losses[pu] 0.1874 0.1256*
4-5 0.01288089 0.00084849 4.832 Reactive power losses[pu] 0.8077 0.6457*
8-5 0.01288089 0.00084849 4.832
5-6 0.01173823 0.00030459 4.832
9-6 0.01173823 0.00030459 4.832 Table 7
Initial operating point for the 25-bus test system.

Bus V [pu] PG[pu] QG[pu] PL[pu] QL[pu]

Table 3
Electrical parameters in pu of the 25-bus test system. Vb = 4.16 kV, Ab = 1 MVA. 1 1.1090 1.4736 1.0669 0 0
2 1.1056 1.4196 0.0754 0 0
Branch R [pu] X [pu] Ix [pu] 3 1.1058 0.4120 1.2513 0 0
1 11 0.004226 0.003111 2.616 Real power losses[pu] 0.0853 0.0800*
11 12 0.004226 0.003111 2.616 Reactive power losses[pu] 0.8077 0.0670*
11 10 0.003232 0.002841 4.021
10 9 0.005387 0.004736 4.021
9 7 0.005387 0.004736 4.021 Table 8
7 16 0.005387 0.004736 4.021 Optimized Operating point for the 9-bus test system by DOPF.
7 14 0.005387 0.004736 4.021
Bus V [pu] PG [pu] QG [pu] Real power Reactive Power
14 15 0.003232 0.002841 4.021
losses losses
14 17 0.006339 0.004667 2.616
7 6 0.005387 0.004736 4.021 1 1.1090 1.7235 0.2563 0.1115 0.5717
6 8 0.010773 0.009471 4.021 2 1.1056 2.0316 0.1177
6 25 0.005387 0.004736 4.021 3 1.1058 0.7064 0.5632
25 19 0.004112 0.007349 6.852 Number of iterations for convergence 3
25 13 0.002056 0.003674 6.852
13 4 0.002056 0.003674 6.852
4 5 0.005387 0.004736 4.021
4 23 0.004309 0.003788 4.021
situation, generation units are operated such as to produce high
23 24 0.004309 0.003788 4.021 power losses and large voltage drops.
24 3 0.008452 0.006223 2.616 In Tables 6 and 7, the real and reactive powers of loads and gen-
13 18 0.005387 0.004736 4.021 erators related to the starting configurations are reported for the
18 21 0.008452 0.006223 2.616
9-bus and 25-bus test systems. In the same table, real and reactive
21 22 0.008452 0.006223 2.616
18 20 0.005387 0.004736 4.021 losses are also given. In Tables 6 and 7, the power losses evaluated
20 2 0.008452 0.006223 2.616 for the initial operating condition using an exact power flow algo-
rithm described in [33] are also reported and indicated with the
symbol *. In this paper, the loads and generators models rely on
Table 4 frequency dependent expressions.
Power limits of generators in pu of the 9-bus test system.
For the solution of the distributed OPF of the 9-bus test, gener-
Bus PGx[pu] PGm [pu] QGx[pu] QGm[pu] ators 1 and 2 are assumed as PV nodes, while the remaining nodes
1 3.5 0.1 0.5 −0.5 are assumed as PQ nodes. Besides, generator 3 is considered as a not
2 3.5 0.1 0.5 −0.5 dispatchable generator (i.e. photovoltaic source). From the analysis
3 3.5 0.1 0.5 −0.5 of bus voltage, it can be noted that at bus 6, Kirchhoff’s law must
be applied. Branch 6-9 is indeed constrained and bus 5 must solve
the optimization problem (Eq. (26)).
Table 5
In Table 8, the DOPF solution is reported with a tolerance equal
Power limits of generators in pu of the 25-bus test system.
to 0.01 for the stopping conditions (Eqs. (44) and (45)) and limit
Bus PGx[pu] PGm[pu] QGx[pu] QGm [pu] values on voltages equal to ±5% of rated voltage (equal to 1 pu).
1 2 0.1 1.5 −1.5 The improvement as compared to the initial operating point is of
2 2 0.1 1.5 −1.5 29.22% in relative percentage value.
3 2 0.1 1.5 −1.5 For the solution of the distributed OPF of the 25-bus test system,
generators 2 and 3 are assumed as PV nodes, while the remaining
nodes are assumed as PQ nodes. Moreover, generator 1 is con-
4. Application sidered as a not dispatchable generator. From the analysis of bus
voltages, it can be observed that the buses 7 and 13 must solve the
Two test systems have been considered for the experiments: linear system (Eq. (26)), because they have got two branches with
a 9-bus system taken from Ref. [30], Fig. 2, and a 25-bus system flows entering the node. After the second iteration in the Backward
taken from Ref. [31], Fig. 5. The electrical data of both systems, the phase, bus 7 must apply the Kirchhoff’s law, because branch 7-9 is
reference base values for voltage (Vb) and power (Ab) and the limits constrained.
of current of each branch are reported in Tables 2 and 3, while the In Table 9, the DOPF solution is reported.
power limits of generators are reported in Tables 4 and 5. Tables 10 and 11 provide a comparison between the results
In Tables 4 and 5 PG and QG are the generated real and reac- attained using the exact load flow algorithm deployed in Ref. [34]
tive power, and subscripts m and x indicate, respectively, their for the OPF solution by Glow-worm Swarm Optimization (GSO) and
minimum and maximum limits. the losses evaluated using the proposed DOPF. This power flow
To perform meaningful tests, an un-optimized starting operat- used in Ref. [34] relies on a methodology developed in Ref. [33]
ing condition is considered for both test systems. According to this using a frequency dependent model both for generators and loads.
282 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283

Table 9 Table 13
Optimized Operating point for the 25-bus test system by DOPF. Comparison between GSO and DOPF for 25-bus test system in terms of relative
improvement.
Bus V [pu] PG [pu] QG [pu] Real power Reactive
losses power losses Bus GSO DOPF

1 1.10900 1.47360 1.06690 0.0667 0.0560 PG [pu] QG [pu] PG [pu] QG [pu]

2 1.10560 0.71308 0.54965
3 1.10580 1.09992 0.72547 1 1.4736 1.0669 1.4736 1.0669
Number of iterations for convergence 4 2 0.8627 0.7794 0.7131 0.5496
3 0.8539 0.5088 1.0999 0.7254
Real power losses (pu) 0.0761* 0.0667
Table 10 Reactive power losses (pu) 0.0634* 0.0560
Comparison between exact power losses evaluation and DOPF power losses evalu- Relative improvement as 10.78% 21.81%
ation for 9 bus test system. compared to initial solution

Bus Exact load flow DOPF

able operating conditions, practically the solution outputted by the
PG [pu] QG [pu] PG [pu] QG [pu]
DOPF can be implemented in G-1 generators, leaving free the G-th
1 1.7227 0.3329 1.7235 0.2563 generator. In this way, the latter would act like a slack bus (i.e. fix-
2 2.0316 0.1177 2.0316 0.1177
3 0.7064 0.5632 0.7064 0.5632
ing the bus voltage and providing frequency regulation) and would
Real power losses (pu) 0.1256* Frequency 0.1115 absorb the errors deriving by a slightly incorrect operating point
Reactive power losses (pu) 0.6457* f = 0.99 0.5717 definition, thus giving rise to a perfectly balanced operation.
(pu) In case more units are droop regulated, these units will absorb
the power unbalance deriving from the approximation errors.
Table 11
Comparison between exact power losses evaluation and DOPF power losses evalu- Convergence properties can be deduced from the literature on
ation for 25-bus test system. Backward/Forward methodologies [35].
The integration of the Backward/Forward approach with the
Bus Exact load flow DOPF
distributed optimization is inherent to the Backward/Forward pro-
PG [pu] QG [pu] PG [pu] QG [pu] cedure which itself is ultimately a depth first visit of a tree network.
1 1.4736 1.0669 1.4736 1.0669 The visit in the considered approach is physical, namely buses
2 0.7132 0.5496 0.7131 0.5496 exchange data at each step and these data are either the output of
3 1.0999 0.6059 1.0999 0.7255 local elaborations or local measures. It could be argued at this point
Real power losses (pu) 0.0800* Frequency 0.0667
Reactive power losses (pu) 0.0670* f = 1.02 0.0560
that the network is not necessarily a tree, since weakly meshed sys-
(pu) tems may also be considered at distribution level. To solve this issue
the authors have elaborated over the concept of tracing according
Table 12 to what suggested by Kirschen et al. in Ref. [36]. The idea underly-
Comparison between GSO and DOPF for 9-bus test system in terms of relative
ing the cited paper was that to consider the domain of a generator
improvement.
as the set of buses that could be reached by the power produced
Bus GSO DOPF by the same generator. In this way, the network was divided into
PG [pu] QG [pu] PG [pu] QG [pu] commons of buses. Once this assumption is made, the main issue
is that to understand how the power is shared among the paths
1 1.7730 0.2092 1.7235 0.2563
2 2.1150 0.2628 2.0316 0.1177 supplying the sink buses. This issue is considered as an optimiza-
3 0.7064 0.5632 0.7064 0.5632 tion problem, which appears to be convex (see Eq. (6)). Therefore,
Real power losses (pu) 0.1227* 0.1115 the problem at hand in each sink bus is a convex problem and the
Reactive power losses (pu) 0.6332* 0.5717 assurance about converge relies on the convergence of a standard
Relative improvement as 21.60% 29.22%
compared to initial solution
Backward/Forward load flow algorithm [35] and of the reinforce-
ment learning algorithms.
Differently from other approaches, as an example [23], the con-
In this case, the generation-load unbalance results in a frequency straints are here considered heuristically, namely by feedbacks or
variation. It must be observed then that, in order to assess the feasi- flags that inhibit or not the optimization of the power flow in a
bility of the operating solutions outputted by the DOPF, an accurate branch. Also the connection among nodes considered by coupling
load flow model described in Ref. [33] has been used to evaluate variables is here created by means of heuristic weights that provide
the power losses. Since optimized operating point may turn to be feedback about the upstream situation. In essence, the proposed
suboptimal and no control over frequency is carried out, also the approach is a backward forward methodology in which at some
frequency in the optimized configuration is reported. buses, entering power flows can be modified based on the new load
Since the two methods use a different load flow algorithm, sharing information outputted by optimization. Such new flows
results have been compared in terms of relative improvement values are then reported backwards to the generators; the latter in
attained by the two methodologies, as reported in Tables 12 and 13. the forward phase recalculate the new power flows. The load flow
In Tables 12 and 13, it can be observed that the improvement of errors deriving from a premature stop of Backward/Forward load
the GSO method compared to the initial operating point is 21.60% flow have been indeed improved using a more compliant conver-
and 10.78%, respectively for the 9-bus and 25-bus test systems. gence condition, see Eq. (44).
Meanwhile, the DOPF solutions carry an improvement (as com-
pared to the initial operating point) of 29.22% and 21.81% in relative 5. Conclusion
percentage value, respectively for the 9-bus and the 25-bus test
systems. In this paper, a new distributed heuristic algorithm for
The results in Tables 10 and 11 show that in both cases, the fre- minimum losses OPF in islanded MGs and, more in general, in dis-
quency remains within bounds, this means that the power losses tribution systems is proposed. The algorithm is easy to understand,
evaluations carried out during the DOPF are a good indicator for fast and reasonably accurate. It requires limited local computa-
power losses minimization. However in order to produce reli- tional requirements and narrow telecommunication bandwidth.
 E. Riva Sanseverino et al. / Electric Power Systems Research 152 (2017) 271–283
The algorithm can be implemented for islanded and meshed micro-
grids. Results show its efficiency as compared to a centralized
approach using a heuristic approach based on Glow-worm Swarm
Optimization.
The algorithm uses heuristic weights to account for the coupling
between nodes belonging to the same area supplied by a single
generator. The weights are learned along the process. To limit the
local computational burden, inequality constraints are considered
through special feedback signals that inhibit the optimization of
the flows in certain branches.
As first full application of this novel approach for DOPF, the
authors have preferred only to consider power losses minimization
and not to include other objectives.
Further direction of this research includes the consideration of
further objectives (production cost, emissions), of possible unbal-
anced operation, which is quite common in low voltage systems,
as well as an experimental assessment on a laboratory test MG.
References
[1] Department of Energy Office of Electricity Delivery and Energy Reliability,
Summary Report: 2012, DOE Microgrid Workshop, 2017, http://energy.gov/
sites/prod/files/. 2012. (Accessed 20 June 2016).
[2] Smart Grids European Technology Platform, http://www.smartgrids.eu/.
(Accessed 20 June 2016).
[3] S.Y. Lin, J.F. Chen, Distributed optimal power flow for smart grid transmission
system with renewable energy sources, Energy 56 (2017) 184–192.
[4] B. Venkatesh, Optimal power flow in radial distribution systems, IPEC, 2010
Conference Proceedings (2010) 18–21.
[5] S.Y. Abdelouadoud, R. Girard, F.P. Neirac, T. Guiot, Optimal power flow of a
distribution system based on increasingly tight cutting planes added to a
second order cone relaxation, Int. J. Electr. Power Energy Syst. 69 (2017) 9–17.
[6] N. Jayasekara, P. Wolfs, M.A.S. Masoum, An optimal management strategy for
distributed storages in distribution networks with high penetrations of PV,
Electr. Power Syst. Res. 116 (2014) 147–157.
[7] E. Riva Sanseverino, N. Nguyen Quang, M.L. Di Silvestre, J.M. Guerrero, C. Li,
Optimal power flow in three-phase islanded microgrids with inverter
interfaced units, Electr. Power Syst. Res. 123 (2015) 48–56.
[8] Q. Xiao, Comparing three methods for solving probabilistic optimal power
flow, Electr. Power Syst. Res. 124 (2015) 92–99.
[9] K. Hassan Youssef, Optimal management of unbalanced smart microgrids for
scheduled and unscheduled multiple transitions between grid-connected and
islanded modes, Electr. Power Syst. Res. 141 (2016) 104–113.
[10] M. Costley, S. Grijalva, Efficient distributed OPF for decentralized power
system operations and electricity markets, 2012 IEEE PES Innovative Smart
Grid Technologies (ISGT) (2012) 1–6.
[11] S. Magnusson, P.C. Weeraddana, C. Fischione, A distributed approach for the
optimal power-flow problem based on ADMM and sequential convex
approximations, IEEE Trans. Control Netw. Syst. 2 (2015) 238–253.
[12] A. Dimeas, N. Hatziargyriou, Operation of a multiagent system for microgrid
control, IEEE Trans. Power Syst. 20 (3) (2005) 1447–1455.
[13] Z. Wang, K. Yang, X. Wang, Privacy-preserving energy scheduling in microgrid
systems, IEEE Trans. Smart Grid 4 (4) (2013) 1810–1820.
[14] A. Dominguez-Garcia, C. Hadjicostis, Distributed algorithms for control of
demand response and distributed energy resources, in: Proc. IEEE CDC,
Orlando, FL, 2011.
[15] Y. Zhang, N. Gatsis, G. Giannakis, Robust energy management for microgrids
with high-penetration renewables, IEEE Trans. Sustain. Energy 4 (4) (2013)
944–953.
[16] E. Crisostomi, M. Liu, M. Raugi, R. Shorten, Plug-and-play distributed
algorithms for optimized power generation in a microgrid, IEEE Trans. Smart
Grid 5 (4) (2014) 2145–2154.
[17] N. Cai, Linearized and Distributed Methods for Power Flow Analysis and
Control in Smart Grids and Microgrids, Order No. 3633944 Michigan State
University, 2014. Ann Arbor: ProQuest.Web. 1, 12/2014.
283
[18] P. Wan, M.D. Lemmon, Optimal power flow in microgrids using
event-triggered optimization, Proceedings of the 2010 American Control
Conference (2010) 2521–2526.
[19] W. Shi, X. Xie, C.C. Chu, R. Gadh, Distributed Optimal Energy Management in
Microgrids, IEEE Trans. Smart Grid 6 (3) (2015) 1137–1146.
[20] S. Bolognani, G. Cavraro, S. Zampieri, A distributed feedback control approach
to the optimal reactive power flow problem, Control of Cyber-Physical
Systems: Workshop held at Johns Hopkins University (2017), March 2013,
D.C. Tarraf, (Ed.) Heidelberg: Springer International Publishing, 2013,
259–277.
[21] P. Šul, S. Backhaus, M. Chertkov, Optimal distributed control of reactive power
via the alternating direction method of multipliers, IEEE Trans. Energy
Convers. 29 (2014) 968–977.
[22] K. Christakou, D. Tomozei, J. Le Boudec, M. Paolone, AC OPF in radial
distribution networks—part I: On the limits of the branch flow convexification
and the alternating direction method of multipliers, Electr. Power Syst. Res.
143 (2017) 438–450.
[23] K. Christakou, D. Tomozei, J. Le Boudec, M. Paolone, AC OPF in radial
distribution networks—part II: an augmented Lagrangian-based OPF
algorithm, distributable via primal decomposition, Electr. Power Syst. Res.
150 (2017) 24–35.
[24] E. Dall’Anese, H. Zhu, G.B. Giannakis, Distributed optimal power flow for
smart microgrids, IEEE Trans. Smart Grids 4 (3) (2014).
[25] Q. Peng, S.H. Low, Distributed optimal power flow algorithm for radial
networks, I: balanced single phase case, IEEE Trans. Smart Grid PP (99) (2017)
1–11.
[26] J. Liu, M. Benosman, A.U. Raghunathan, Consensus-based distributed optimal
power flow algorithm, in: Innovative Smart Grid Technologies Conference
(ISGT), 2015, IEEE Power & Energy Society, Washington, DC, 2015, pp. 1–5.
[27] E. Riva Sanseverino, N.Q. Nguyen, M.L. Di Silvestre, R. Badalamenti, G. Zizzo,
Microgrids Optimal Power Flow Through Centralized and Distributed
Algorithms, 2017, https://workshopcleaned.sciencesconf.org/file/189174.
Hanoi, 12/2015. (Accessed 20 June 2016).
[28] A. Khodaei, M. Shahidehpour, J. Choi, Optimal hourly scheduling of
community-aggregated electricity consumption, J. Electr. Eng. Technol. 8
(2013) 724–733.
[29] M. AboGaleela, M. El-Sobki, M. El-Marsafawy, A two level optimal DSM load
shifting formulationusing genetics algorithm case study: residential loads,
Proc. IEEE Power Africa Power Engineering Society Conference and Exposition
in Africa (2012) 1–7.
[30] A.H. Mohsenian-Rad, A. Leon-Garcia, Optimal residential load control with
price prediction in real-time electricity pricing environments, IEEE Trans.
Smart Grid 1 (2) (2010) 120–133.
[31] E. Riva Sanseverino, M.L. Di Silvestre, R. Badalamenti, N.Q. Nguyen, J.M.
Guerrero, L. Meng, Optimal power flow in islanded microgrids using a simple
distributed algorithm, Energies 8 (10) (2015) 11493–11514.
[32] G.K. Raju, P.R. Bijwe, Efficient reconfiguration of balanced and unbalanced
distribution systems for loss minimization, IET Gen. Transm. Distrib. (2008)
7–12.
[33] E. Riva Sanseverino, N.Q. Nguyen, M.L. Di Silvestre, G. Zizzo, J.M. Guerrero, A
parametric study on unbalanced three phase islanded microgrids with
inverter interfaced units, in: Proceedings of the 2013 AEIT Annual Conference,
IEEE Press, 2013.
[34] E. Riva Sanseverino, N.Q. Nguyen, M.L. Di Silvestre, F. de Bosio, Q.T.T. Tran, G.
Zizzo, Frequency constrained optimal power flow based on glow-worm
swarm optimization in islanded microgrids, 2015 AEIT International Annual
Conference (AEIT) (2015) 1–6.
[35] E. Bompard, E. Carpaneto, G. Chicco, R. Napoli, Convergence of the
backward/forward sweep method for the load-flow analysis of radial
distribution systems, Int. J. Electr. Power Energy Syst. 22 (2000) 521–530.
[36] D. Kirschen, R. Allan, G. Strbac, Contributions of individual generators to loads
and flows, IEEE Trans. Power Syst. 12 (1) (1997) 52–60.
[37] N. Cesa Bianchi, G. Lugosi, Prediction, Learning, and Games, Cambridge
University Press, 2016.
[38] J.M. Guerrero, J.C. Vasquez, J. Matas, L.G. De Vicuna, M. Castilla, Herarchical
control of droop-controlled AC and DC microgrids – a general approach
toward standardization, IEEE Trans. Ind. Electron. 58 (2011) 158–172.
[39] E. Riva Sanseverino, M.L. Di Silvestre, G. Zizzo, G. Graditi, Energy efficient
operation in smart grid: optimal management of shiftable loads and storage
systems, Proc. Int. IEEE Symposium on Power Electronics, Electrical Devise,
Automation and Motion (2012).