Electric Power Systems Research 171 (2019) 230–239

Contents lists available at ScienceDirect

Electric Power Systems Research

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

A multiagent design for self-healing in electric power distribution systems T

a,b a,b,⁎
E. Shirazi , S. Jadid
a
Electrical and Electronic Engineering School, Iran University of Science and Technology, Tehran, Iran
b
Green Research Center, Iran University of Science and Technology, Tehran, Iran

ARTICLE INFO ABSTRACT

Keywords: The out-of-service area after fault should berestored by changing the distribution system configuration by means
Distributed control of switching actions on the feeders. This paper proposes an agent-based approach for service restoration in smart
Multi agent systems distribution systems with distributed generations (DG). The proposed multi agent has four different types of
Self-healing agents: feeder agents, zone agents, switch agents and DG agents. The agents can communicate and cooperate
Service restoration
with each other in order to supply services to out-of-service customers. An artificial neural network has been
Smart grid
Distributed generation
considered within DG agent to predict the DG generation. The restoration plan is built based on local data
Neural network considering system conditions, operational constraints and fault location. Different loading conditions have been
considered under different scenarios and the result of proposed multi agent approach for each scenario with and
without DG have been compared. Simulation results show the efficiency of proposed agent architecture.

1. Introduction communication. Local data is acquired via local sensors, processed lo-
cally without the need of any SCADA-based central station. The whole
The introduction of distributed generation (DG) brings fundamental system behavior is sum of individual local actions [7].
and fast changes to power systems. The purpose of incorporating re- Multi agent systems (MAS) have come to the light as a liable tech-
newable energy such as photovoltaic (PV) system on the smart grid is to nology to apply distributed control strategies in power systems [8]. A
increase system reliability, reduce unrestored areas after fault in dis- MAS based self-healing approach for urban power grid which has five
tribution systems and improve system conditions as well as environ- operating states and four sub controls, was presented in Ref. [9].
mental and energy concerns. The rapid growth in integrating PV system Agent-oriented designs offer multiple pros in Smart Grid simulations
can cause problems if not managed properly. [10]. The MAS can solve problems much faster [11].
Self-healing is the ability of distribution systems to automatically A multi agent system has been proposed in Ref. [12] to solve re-
restore themselves after a permanent faults. It is the key characteristic storation problem of distribution network in distributed manner. The
of a smart grid by the report from National Energy Technology application of a MAS for improving voltage profile and DG optimal
Laboratory (NETL) of the United States [1]. Service restoration aims to dispatch was proposed in Ref. [13]. A set of relay agents which com-
restore loads after a fault by changing system topology while meeting municate and cooperate to apply an adaptive current differential re-
operational constraints [2]. Various studies have been done to address laying function was proposed in Ref. [14]. In order to implement self-
the service restoration problem, in either centralized or decentralized healing, the agents have to adjust to the conditions of the system. A
approaches, which having their own pros and cons [3] (i.e. heuristic MAS to restore outage area after fault in a distribution system was
algorithms [4], mathematical programing [5], fuzzy logic [6], etc.). proposed in Ref. [15]. Although the load priorities, load shedding or
The main advantage of centralized approaches is that they can ob- increase backup capacity by load transfer from some feeders to the
tain the best solution to the problem. In these approaches, in order to other ones had not been considered. An agent-oriented energy man-
obtain a solution a central controller (CC) process the gathered data agement system was proposed in Ref. [16], which control DGs in dis-
from whole system. Therefore, they are prone to be affected by a single tribution systems in order to optimize system operation and efficiency.
point of failure (SPOF). They also impose the challenge of managing big An adaptable agent based model is described in Ref. [17] to simulate
data and have to support heavy computing processes. Decentralized the dynamic of smart city. A MAS model for the self-healing protection
approaches are more common for large networks and mostly work on system was proposed in Ref. [18].
parallelizing the problem. These approaches are based on peer-to-peer In Ref. [19], a multi-agent based system by help of unified data

⁎
Corresponding author at: Electrical and Electronic Engineering School, Iran University of Science and Technology, Tehran, Iran.
E-mail address: [email protected] (S. Jadid).

https://doi.org/10.1016/j.epsr.2019.02.025
Received 16 September 2018; Received in revised form 14 February 2019; Accepted 22 February 2019
0378-7796/ © 2019 Elsevier B.V. All rights reserved.
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

Nomenclature Di The demand of ith zone

Si The status of ith zone
Abbreviations Pi The load priority of the ith zone
Sitie The status of tie switch ith
ANN Artificial neural network S CB
j The status of CB jth
BDI Believe-desire-intension Iz The available current in zone z
BFA Backup feeder agent Vz The available voltage in zone z
CC Central controller Z fo The series impedance of path between node with
DGA Dg agent minimum voltage and substation
EDLC Electric double-layer capacitor I feeder The maximum additional load of feeder
ESS Energy storage system SCBF The spare capacity of each backup/supporting feeder
FA Feeder agent IPV Light generated current of a PV cell
GR Group restoration ISC Short circuit current of PV cell
HSA Head switch agent G Sun irradiance
MAS Multi agent systems KI The short-circuit current/temperature coefficient
MFA Main feeder agent T Environment temperature
MLR Maximum load restoration Reference temperature
MSC Minimum switching cost Id The diode current of the ideal PV cell
NAR Nonlinear autoregressive I0 The diode ideality constant
NETL National energy technology laboratory a The equivalent series resistance of the module
PF Power flow RS The equivalent parallel resistance
PR Partial restoration RP The equivalent parallel resistance
PV Photovoltaic Vt The thermal voltage of the module
SA Switch agent k The Boltzmann constant
SSA Sectionalizer switch agent q Electron charge
SFA Supporting feeder agent PSolar The output of PV stations
SPOF Single point of failure
Max
PESS The maximum output power of ESS
SR Sectional restoration PDG The power each DG station can deliver
TSA Tie switch agents PNwk The available capacity of feeder to carry out DG output
ZA Zone agents
Path
Imax The minimum of current upper bound of path between DG
ZBM Zinc bromide battery modules and respective tie switch
f (w ) The loss function of neural network
Parameters and Variables J Jacobian matrix
E Errors of loss function
Ij Currents in branch j Damping factor
IMax Maximum allowable current I The identity matrix
Vi Voltages of nodes i H Hessian matrix
Vmin Minimum allowable voltage W Weight vector
VMax Maximum allowable voltage

platform for fault management was proposed. A MAS with fuzzy clus- considering load priority, DG uncertainties and proper control and
tering algorithm was proposed in Ref. [20]. In Ref. [21] an agent-based communication structure has been presented in this paper. The rest of
platform for reconfiguration and fault diagnosis with high-level archi- the paper is organized as follows. Section 2 formulates the restoration
tecture was developed. In order to coordinate the operation of relays problem. The PV generation forecast is presented in Section 3. The
during faults, the distance between the sample data and the cluster proposed multi agent system architecture is described in Section 4. The
center had been used. An autonomous MAS to control switching action simulation result presented in Section 4 and finally Section 5 concludes
is proposed in Refs. [22,23]. the paper.
BDI (Believe–Desire–Intension)-based multi-agent architecture is
proposed in Ref. [24]. The multiagent based load restoration algorithm 2. Restoration problem formulation
based on Knapsack problem was presented in Ref. [25]. A fuzzy-mul-
tiagent self-healing scheme has been proposed in Ref. [26]. A fuzzy In this paper, the self-healing problem in a smart distribution net-
rule-based system is employed for decision-making support in MAS but work is modeled as a multiobjective problem. The objectives of re-
it does not consider any topology related issue. storation can be maximizing of restored load and/or minimizing the
These studies did not consider an appropriate communication and switching costs, number of switching. Many studies considered loss
control structure for each agent set. Furthermore, they did not consider minimization during restoration process, but it is believed that when
DGs and the challenges involved with DG integration. The main issue to system is in emergency mode, losses are not a matter of importance.
be considered is uncertainty of PVs and fluctuation of output power due The constraints of the service restoration problem are voltage and
to intermittent behavior of sun irradiance and temperature. To reduce current limits and maintaining radial topology of distribution network.
the impact of fluctuations an appropriate prediction method has been The currents in all branches should be less than maximum allowable
used in this study, also an energy storage system considered along with current:
PV to smooth out the output. On the other hand, a precise modelling of
PV module should be considered to design a PV system for application Ij IMax (1)
in practical systems.
A well-designed multi agent framework for service restoration The voltages of nodes should be kept within their limits, which are
0.9 and 1.05 P.U. based on ANSI C84.1 standard:

231
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

Vmin Vi VMax (2) where out-of-service area are restored in turn, section by section and by
The first objective, maximum load restoration (MLR) is to restore as number of switching (whether closing a tie switch or opening a CB),
many loads as possible considering loads priorities: and third is partial restoration (PR), where some out-of-service area
cannot be restored due to lack of spare capacity or violation of con-
straints and some loads have to be shed.
MLR = Max Di × Pi × Si
i (3)
3. Proposed MAS architecture
The outage area consists of multiple zones. If there are N zones in
the outage area, then there are 2N combinations of zones to be restored. The main goal of self-healing in this paper is to increase reliability of
Some of these combinations for restoration are feasible and some of distribution system, minimize out-of-service area after a fault by using
them are infeasible. After omitting the infeasible combination, then the distributed control approaches. In order to make good decision, it is
MLR for each feasible combination is calculated based on Eq. (3). The crucial to gather enough information. Thus for multi-agent based re-
Di is the demand of ith zone, Pi is the priority of ith zone, Si is a binary storation, communication is a key technology. The communication in-
variable indicating the status of ith zone, if the ith zone is in the frastructure is independent of the physical structure of the smart dis-
combination then it will be one, otherwise it will be zero. One of the tribution network. Hence, even if there is a fault in distribution line, the
objectives is to restore as many loads as possible considering their communications between two agents can still be persistent [25]. In this
priorities, so the combination with the most MLR will be considered for study, it has been assumed that the communication is reliable and will
restoration, regarding operational constraints. not encounter any problem. The communications among various agents
The second objective, minimum switching cost (MSC), is mini- are shown in Fig. 1.
mizing the number of switching considering both ties and circuit In this study, it is also assumed that the location of fault is defined so
breakers. fault location algorithm has been skipped. Only single fault cases have
been considered for all of the systems.
MSC = Min Sitie + (1 S CB
j )
Ties CBs (4) 3.1. Agent types
where and
Sitie are status of tie switch ith and CB jth respectively. The
S CB
There are four types of agents in the proposed architecture:
j
overall objective of restoration problem can be defined as follow:

OBJ = MLRMSC = Max( D × Pi × Si ) + Min( Sitie + (1 S CB

j )) 1 Feeder Agents (FA) are responsible for energizing the respective
i i Ties CBs feeder. They process the data from other agents and make decisions
(5) based on local data it gathers to restore loads. There are three types
Three different types of restoration have been considered in this of feeder agents: Main Feeder Agent (MFA) the feeder agent of faulty
study. First group restoration (GR), where all out-of-service area are feeder, Back-up Feeder Agent (BFA): the feeder agents of neigh-
restored by closing one tie switch. Second, sectional restoration (SR), boring feeder of faulty feeder and Supporting Feeder Agent (SFA):

Fig. 1. Communications among agents.

232
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

the feeder agents of neighboring feeder of back-up feeders when the BFA sends request messages to its SFAs and asks for SC fol-
2 Switch Agents (SA) distinguish the out-of-service area, all switches lowing operation has to be done in order to calculate spare capacity of
within the outage area and the topology of it. There are three types each backup/supporting feeder:
of switch agents: Head Switch Agent (HSA): the downstream switch The BFA/SFA sends a request message to its ZAs and asks for current
agent of faulty section that is responsible for defining topology of and voltage data to calculate maximum load they can supply.
healthy out-of-service zones, Sectionalizer Switch Agents (SSA) and Step 1: Each ZA replies with an inform message containing lowest
Tie Switch Agents (TSA). voltage of its nodes and minimum difference between the current upper
3 Zone Agents (ZA) are responsible for calculating remaining capacity, bound and its branches current before the fault.
voltage and current of zones. The ZAs of out-of-service area are
I z = min(Imax z
Ibranch ) (6)
responsible for declaring the demand.
4 DG Agent (DGA) are responsible for calculating and declaring the z
V z = min(Vnode ) (7)
power, which can be provided by respective DG.
Step 2: The BFA/SFA receives zones data and calculates SC by using
The agents and their task will be described thoroughly in following following equations [27]:
section. min(V z ) Vmin
When a fault occurs in a feeder, the SSA of upstream circuit breaker I fv =
Z fo (8)
(CB) opens and sends a message to the FA of respective feeder and
declares the occurrence of a fault. After isolation the fault, the re- I fc = min(I z ) (9)
storation process begin. Then this SSA called HSA and this FA called
MFA, which stand for head switch and main feeder agents, respectively. Where Vmin is minimum allowable voltage in network and is the Z fo
series impedance of path between node with minimum voltage and
substation. The shunt admittance in branches are neglected.
3.2. Strategy and tasks of agents Step 3: The maximum additional load, without violation of current
and voltage constraints in the backup feeder can be obtained by:
Agents are designed to energize out of service zones and make de-
cisions based on local data. The pseudocode of the restoration process is I feeder = min(I fc , I fv ) (10)
shown in Fig. 2.
The spare capacity of each backup/supporting feeder can be cal-
culated as follow:
3.2.1. The procedure of calculating spare capacity SCBF = Ifeeder × Vmin (11)
When the MFA sends request messages to BFAs and asks for SC, or

Fig. 2. Pseudocode of proposed restoration process.

233
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

Step 4: If BFA receives a request for additional capacity, it starts IPV = ISC . G
. (1 + KI . (T Tref )) (13)
1000
load-transfer process.
Step 5: The BFA sends a request message to its SFA (if available) and where IPV is light generated current of a PV cell, ISC is short circuit
asks for more spare capacity. current of PV cell, G is sun irradiance, KI is the short-circuit current/
Step 6: When the BFA receives an inform message from SFA, it can temperature coefficient, T is environment temperature in Kelvin [K]
define new SC as following: and is reference temperature which is 298.5.
V + RS I
(12)
new
(14)
old, rem
SCBF = min(D out SCBF , SC SF ) Id = I0 (e a . Vt 1)

Step 7: Where SCBF

new
is new SC of backup feeder, SCBF
old, rem
is remaining where Id is the diode current of the ideal PV cell, I0 is the diode sa-
SC of backup feeder, D out is demand of remaining unrestored zones and turation current, a is the diode ideality constant.
SCSF is SC of supporting feeder. V + RS I V + RS I
I = IPV . NP I0. NP . (e a . Vt 1)
RP (15)
3.2.2. The procedure of topology determination
RS is the equivalent series resistance of the module and RP is the
When an occurrence of a fault is detected, the respective switches
equivalent parallel resistance, Vt is the thermal voltage of the module
are opened to isolate the fault. The healthy part of outage area can be
which is given by equation
rstored before clearing the fault which is called outage area or out-of-
service zone in the rest of the paper. NS . k. T
Vt =
Step 1: The MFA sends a request message to HSA to define topology q (16)
of outage area.
where k is the Boltzmann constant [J/K] and q is electron charge [C]
Step 2: Then the HSA starts defining topology of out-of-service zones
and available paths to restore them. cell
V + RS I V + RS I
PPV = V . (I pv . NP I0. NP . (e a . Vt 1) )
Step 3: The HSA sends a request message to downstream SSA and ask RP (17)
about its status. The status of a switch can be either “open” or “close”. If
In order to smooth out PV power fluctuations various types of en-
there are more than one downstream switch the request message passes
ergy storage technologies have been proposed such as electric double-
to all of them. In this case, an inform message will pass to HSA to inform
layer capacitor (EDLC) [28] and battery energy storage [29]. In this
identification of new branch, so HSA becomes aware of number of
study, zinc bromide battery modules (ZBM) has been considered to
branches at its downstream side.
deliver uninterrupted power, efficiently. The energy of ESS at time t is
Step 4: The downstream SAs send topology information to HSA. If
calculated as follows:
the status of a switch is “close”, it passes the request to its downstream
SSA(s) and continue the procedure. EESS (t ) = EESS (t Ch
1) + t × PESS Dch
t × PESS (18)
Step 5: This procedure continues until it reaches a tie switch or
feeder terminal. Then it will start from another point left behind. Since t = 1 we will have:
Step 6: When all outage area has been searched and available paths EESS (t ) = EESS (t Ch
1) + PESS Dch
PESS (19)
have been defined, the HSA sends this information including topology
of outage area and available paths to MFA. The energy storage is charged through PV and discharged if the
Whether the group, sectional or partial restoration execute in the solar station is connected to the network, hence:
system, there are two lists of switches. CBs to be opened and ties to be EESS (t ) = EESS (t 1) + PPV Dch
PESS (20)
closed. The MFA sends switches on this list command messages and ask
them to open/close. Which is the energy of storage in solar station at time t. The power
Step 8: When a SSA receives “open” command, it should acknowl- output of solar station considering the maximum rated power of ESS at
edge and open the respective switch. which the system can be discharged is calculated as follows:
Step 9: When a TSA receives “close” command, it should acknowl- Max
PSolar = min(PESS , PESS ) (21)
edge and close the respective tie switch.
The power each DG station can deliver considering operational
constraints of the network is calculated as follows:
3.2.3. The calculation of DG output
The use of renewable energies has been increased considerably, in PDG = min(PSolar , PNwk ) (22)
order to diminish the effects of global warming and energy crisis. where PNwk is the available capacity of feeder to carry out DG output.
Photovoltaic (PV) gets lots of attentions among all potential renewable
Path
energy resources. PV integration with power grid requires the ability to PNwk = Vmin × (Imax Ifeeder ) (23)
deal with the uncertainty of power output. An accurate forecasting is a
where is the minimum of current upper bound of path between DG
Path
Imax
key factor to handle uncertainty, which provides the grid operators
and respective tie switch.
with significant information to manage the demand and supply of
power grid.
4. Simulation results
A method for forecasting PV generation should handle non-linearity
due to dynamic and non-linear behavior of meteorological data.
To demonstrate the effectiveness of the proposed method, it has
Artificial neural network (ANN) is an appropriate method in compar-
ison to other statistical ones. The forecasting process is brought in
Appendix A.
If a DG Agent (DGA) receives a request from its FA, it declares the
power it can provide. The DGA has the result of ANN and it can cal-
culate the output of DG by environmental factors, PV model and time of
the day.
The solar modules can be modeled as shown in Fig. 3. There are Ns
cells connected in series and Np cells of parallel connections in each
module Fig. 3. The equivalent circuit of PV module.

234
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

been applied to a distribution system including 4 feeders, 70 nodes and distribution systems presented by Ref. [32] has been used. The PF has
78 branches [30] and shown in Fig. 4. Before the fault, all CBs are been coded in Matlab and a Matlab–Netlogo extension has been used
normally closed and all ties are normally opened and tie lines are shown for this purpose [33].
by dash lines. The output power of each PV module and energy stored in ESS
In this study, BP SX 150 solar module has been considered, which is without any discharge are shown in Fig. 5
made of 72 multi-crystalline silicon solar cells in series. A Raytheon When a fault occurs, the respective switched will be opened to
Ktech’s 30 kW energy storage system (ESS) has been used to manage PV isolate the fault, and then the downstream feeder will be out of service.
ramp rates. The 30 kW/120 kW h ESS has 12 RedFlow® ZBM and de- After isolating the fault, the HSA will sends a signal to MFA. After that,
signed to deliver uninterrupted power, efficiently. the HSA tries to find possible paths for restoration and at the same time,
A NARNN has been used to predict sun irradiation and environment MFA talks with BFAs and asks for available capacities. Then the MFA
temperature. It has 10 hidden layers and two delays. runs PF for available paths. If it ran successfully the MFA keeps the
There are three DGs in the network, which are shown in the Fig. 4. combination, otherwise the combination will be omitted.
Each DG consists of a solar station and an ESS. There are 10 modules in In this study, it has been assumed that the location of fault is defined
each solar array and 10 arrays in each solar station. so fault location algorithm has been skipped. Only single fault cases
The proposed algorithm has been applied to different scenarios. In have been considered in this paper. There are many possibilities for
the first scenario, system is in off-peak mode and/or the demand of fault location. Some of possibilities are shown in Fig. 6. The worst case
outage area is less than the capacity of at least one backup feeder, so the for fault to happen is where all of loads in the feeder is out-of-service
group restoration is possible. In the second scenario, the demand of and should be restored as shown in Fig. 6(a).
outage area is more than maximum capacity of backup feeders and the The loading data, which is used for defining loading factor, is based
summation of all backup capacities is enough to restore all loads, hence on real data from Jackson Associates’ MAISY Utility Customer Energy
the sectional restoration should be done. In the third scenario, the de- Use and Hourly Loads Databases [34]. Then by applying an ANN,
mand of outage area is more than maximum capacity of backup feeders loading factor of the network for different hours based on one-year data
but the summation of all backup capacities is not enough to restore all has been calculated. Even in the peak hours and the worst case, the load
of loads, so load transfer from supporting feeders should be done. In will not exceed 115% of average load, but in order to show the effi-
fourth scenario, not all capacities from back up and supporting feeder ciency of proposed method, it has been applied on different loading
are enough for load restoration and therefore some of loads should be conditions varying from 100% to 210%, which is unlikely to happen.
shed. The required switching to restore out-of-service zones are shown in
The occurrence of each scenario depends on peak hours and demand Table 1 for each scenario. There are four possibilities for supply and
of outage area. If the distribution system is in peak hours the possibility demand:
of load shedding is high, but during off-peak hour, load shedding is
unlikely to happen. • If the demand of outage area is less than maximum capacity of any
In order to implement proposed multi agent systems, Netlogo [31] backup feeder, then the group restoration is possible. As can be seen
environment has been used. In order to check the feasibility of each in Table 1, the restoration can be done by closing just one tie switch.
scenario, the feeder agent has to run a power flow (PF) for that sce- • If the maximum capacity of backup feeders is less than demand of
nario. In this study the fast-decoupled power flow algorithm for radial outage area and hence the group restoration is not possible and

Fig. 4. Under study 70-bus distribution system.

235
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

Fig. 5. The output power of PV module and energy of ESS.

Fig. 6. Different fault location and respective healthy outage area.

Table 1 The restored areas along with number of switching for scenarios
The simulation resultfor different scenarios. with and without DGs are shown in Fig. 7.
Scenario Loading condition Tie switch #number CBs #number
As can be seen, when the loading increases, the percentage of re-
stored load decreases due to less spare capacity by backup and sup-
1 100% #6 – porting feeders as well as more demand of outage area.
2 130% #3, 7, 9 #8, 10, 11 Furthermore, as can be seen in Fig. 7 the number of switching in-
3 160% #4, 6, 8, 9 #5, 7, 8, 11
creases as the loading increases but at a specific point (after 160%),
4 190% #6, 8 #5, 7
5 210% – – when there is not enough capacity to restore load it will reduce till a
loading condition where 0 % of outage area restored due to lack of
capacity (210% loading).
sectional restoration has to be done.
• If the capacity of all backup feeders cannot meet demand of out-of- 5. Conclusion
service area then the load transfer has to be done.
• If after load transfer and partial restoration, there are still some In this paper, a multi agent system to implement self-healing in
unrestored zone, they have to be shed. smart distribution network has been designed. The proposed control

236
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

Fig. 7. Comparison of restored load and number of switching in different scenarios.

structure consists of three main types of controllers: feeder, switch and scenarios have been analyzed and compared. This scheme worked ef-
zone. The operating strategy and tasks of each controller has been de- ficiently by representing how fast load restoration could be done with
signed regarding the multiagent system concept. Operational con- distributed control by means of local data. Moreover, simulation results
straints of system such as voltage and current limits in the agents de- evidenced that the proposed MAS model offers a feasible solution to
sign. In order to achieve cooperation among various agents, a two-way optimal service restoration problem in smart distribution network. With
communication has been implemented. To show the capability of the the proposed multi agent frame work an autonomous and effective
proposed model, it has been applied on a sample distribution network service restoration can be obtained.
and a simulation model was developed. The results of applying the

Appendix A

Training is a basic operation of an ANN, which tries to figure a relationship between data set and target by updating the weight vector. The
representation of a nonlinear autoregressive (NAR) NN can be seen in Fig. A1, where N is a neuron of ANN.
The training of neural network is the process of learning. The learning problem can be formulated as minimization of a loss function, which is
consist of errors and regulation terms. These terms are used for evaluation of fitting ANN to data and preventing overfitting, respectively.
The loss function depends on the biases and synaptic weights in the neural network considered as a single weight vector w. The W* is the point
where loss function takes minimum. To determine the minimum of loss function, the gradient vector and Hessian matrix should be calculated which
are first and second derivatives respectively.
df
if (w ) = (i = 1, ..., n) = 0, ..., m
dwi (A.1)

d 2f
Hi, j f (w ) = (i, j = 1, ..., n)
dwi. dwj (A.2)

Fig. A1. Representation of ANN.

237
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

Fig. A2. Minimization of loss function in ANN.

In this study, the Levenberg-Marquardt (LM) algorithm has been used to calculate the minimum of loss function. Consider a loss function which
can be expressed as a sum of squared errors

f= ei2 , i = 0, ..., m (A.3)

where m is the number of elements in the data set (Fig. A2).
The gradient vector of loss function is defined as:
f = 2J T . e (A.4)
where the J is Jacobian matrix and can be defined as derivatives of errors:
dei
Ji, j f (w ) = (i = 1, ..., m&j = 1, ..., n)
dwj (A.5)
where m and n are number of elements in data set and parameters of ANN, respectively. The Hessian matrix can be approximated as:
H 2J T . J + I (A.6)
where is damping factor and I is the identity matrix. The damping factor is used to ensure the positiveness of the H. At the beginning, it is large so it
will lead to small steps updates. The damping factor will decrease as the loss decreases. When λ is large, the method takes a small step in the gradient
direction. As the method nears a solution, λ is chosen to be small and the method converges quickly via the Gauss–Newton method. The basic
strategy behind choosing the damping term uses the observations that the square of the step size Δ2 = δwT·δw is a monotonically decreasing function
of λ. Therefore, for a sufficiently large value of λ, the algorithm will take an arbitrarily small step in a descent direction. If a proposed step is
unacceptable, one need only increase the damping term until a smaller, more acceptable step has been found. Because choosing λ is equivalent to
choosing the step size, the Levenberg-Marquardt method can be considered a trust-region method. There are two broad classes of methods for
determining the appropriate damping. This can be done by either adjusting λ directly, or, by first choosing an acceptable step size Δ and then finding
a λ such that |δw| ≤ Δ. We will refer to these two types of schemes as direct and indirect methods respectively. Many schemes have been developed
to efficiently adjust λ or Δ.
The simple method originally suggested by Marquardt is usually adequate. In this scheme, if a step is accepted, then λ is decreased by a fixed
factor, say 10. If a step is rejected then λ is appropriately raised by a factor of 10. The qualitative effect of the damping term is to modify the
eigenvalues of the matrix JTJ + λI to be at least λ. Often, the eigenvalues of JTJ are well spaced on a log-scale; it is therefore natural to choose the
factor by which λ is either raised/lowered to be comparable to the eigenvalue spacing of JTJ.
The LM algorithm defines the parameters improvement process as follows:
wi + 1 = wi (JiT . Ji + iI)
1. (2JiT . ei ), i = 0, 1, ... (A.7)
The first step of LM algorithm is to calculate the loss function, the gradient vector and the Hessian matrix approximation. Then the damping
factor is tuned to reduce errors at each iteration.
The LM algorithm is a very fast algorithm to train ANN in forms of sum of squared errors.

References Energy Society General Meeting (2012) 1–1.

[5] A. Golshani, W. Sun, Q. Zhou, Q.P. Zheng, J. Tong, Two-stage adaptive restoration
decision support system for a self-healing power grid, EEE Trans. Ind. Inf. 13
[1] National Energy Technology Laboratory, A Systems View of the Modern Grid, (December (6)) (2017) 2802–2812.
Available: (2007) https://www.netl.doe.gov/File%20Library/research/energy [6] S.J. Chen, T.S. Zhan, C.H. Huang, J.L. Chen, C.H. Lin, Nontechnical loss and outage
%20efficiency/smart%20grid/whitepapers/ASystemsViewoftheModernGrid_Final_ detection using fractional-order self-synchronization error-based fuzzy petri nets in
v2_0.pdf. micro-distribution systems, IEEE Trans. Smart Grid 6 (1) (2015) 411–420.
[2] P.L. Cavalcante, et al., Centralized self-healing scheme for electrical distribution [7] A. Zidan, et al., Fault detection, isolation, and service restoration in distribution
systems, IEEE Trans. Smart Grid 7 (January (1)) (2016) 145–155. systems: state-of-the-art and future trends, IEEE Trans. Smart Grid 8 (September
[3] P. Parikh, I. Voloh, M. Mahony, Fault location, isolation, and service restoration (5)) (2017) 2170–2185.
(FLISR) technique using IEC 61850 GOOSE, 2013 IEEE Power & Energy Society [8] I.H. Lim, et al., Design and implementation of multiagent-based distributed re-
General Meeting (2013) 1–6. storation system in DAS, IEEE Trans. Power Deliv 28 (April (2)) (2013) 585–593.
[4] M. Kleinberg, K.M. Miller, Improving service restoration of power distribution [9] H. Liu, X. Chen, K. Yu, Y. Hou, The control and analysis of self-healing urban power
systems through load curtailment of in-service customers, 2012 IEEE Power and grid, IEEE Trans. Smart Grid 3 (September (3)) (2012) 1119–1129.

238
E. Shirazi and S. Jadid Electric Power Systems Research 171 (2019) 230–239

[10] S. Boughosn, P. Ranganathan, S. Salem, J. Tang, D. Loegering, K. Nygard, Agent- (2015) 252–259.
oriented designs for a self healing smart grid, 1st IEEE International Conference on [23] N.I. Elkalashy, M. Lehtonen, Decentralized earth fault selectivity using transient
Smart Grid Communications (2010). front slopes for unearthed mv networks, Electr. Power System Res. 63 (2014)
[11] J. Wu, T. Lee, C. Lu, S. Su, An autonomous decision approach for fault allocation 908–916.
and service restoration in electrical distribution systems by multi agent system, [24] Y.T. Pan, M.S. Tsai, Development a BDI-based intelligent agent architecture for
Proceeding of 9th International Conference on Hybrid Intelligent Systems (2009) distribution systems restoration planning, 2009 15th International Conference on
89–94. Intelligent System Applications to Power Systems (2009) 1–6.
[12] J.M. Solanki, S. Khushalani, N.N. Schulz, A multi-agent solution to distribution [25] X. Yinliang, L. Wenxin, Novel multiagent based load restoration algorithm for mi-
systems restoration, IEEE Trans. Power Syst. 22 (2007) 1026–1034. crogrids, IEEE Trans. Smart Grid 2 (March (1)) (2011) 152–161.
[13] M. Baran, I. El-Markabi, A multiagent-based dispatching scheme for distributed [26] A. Elmitwally, M. Elsaid, M. Elgamal, Z. Chen, A fuzzy-multiagent self-healing
generators for voltage support on distribution feeders, IEEE Trans. Power Syst. 22 scheme for a distribution system with distributed generations, IEEE Trans. Power
(February (1)) (2007) 52–59. Syst. 30 (September (5)) (2015) 2612–2622.
[14] Y. Tomita, C. Fukui, H. Kudo, J. Koda, K. Yabe, A cooperative protection system [27] R.M. Ciric, D.S. Popovic, Multi-objective distribution network restoration using
with an agent model, IEEE Trans. Power Deliv. 13 (October (4)) (1998) 1060–1066. heuristic approach and mix integer programming method, Int. J. Electr. Power
[15] M. Tsai, Y. Pan, Application of BDI-based intelligent multi-agent systems for dis- Energy Syst. 22 (October (7)) (2000) 497–505.
tribution system service restoration planning, Euro. Trans. Electr. Power (2011) [28] N. Kakimoto, H. Satoh, S. Takayama, K. Nakamura, Ramp-rate control of photo-
1783–1801. voltaic generator with electric double-layer capacitor, IEEE Trans. Energy Convers.
[16] F. Ren, M. Zhang, D. Sutanto, A multi-agent solution to distribution system man- 24 (June (2)) (2009) 465–473.
agement by considering distributed generators, IEEE Trans. Power Syst. 28 (May [29] J. Traube, L. Fenglong, D. Maksimovic, J. Mossoba, M. Kromer, P. Faill, S. Katz,
(2)) (2013) 1442–1451. B. Borowy, S. Nichols, L. Casey, Mitigation of solar irradiance intermittency in
[17] S. Karnouskos, T. Nass de Holanda, Simulation of a smart grid city with software photovoltaic power systems with integrated electricvehicle charging functionality,
agents, European Modeling Symposium EMS (2009). IEEE Trans. Power Electron. 28 (June (6)) (2013) 3058–3067.
[18] S. Sheng, K. Li, W. Chan, Z. Xiangjun, D. Xianzhong, Agent-based self-healing [30] D. Das, A fuzzy multiobjective approach for network reconfiguration of distribution
protection system, IEEE Trans. Power Deliv. 21 (April (2)) (2006) 610–618. systems, IEEE Trans. Power Deliv. 21 (January (1)) (2006) 202–209.
[19] Y. Liu, P. Sun, C. Wang, Group decision support system for backbone-network re- [31] U. Wilensky, (1999). NetLogo. http://ccl.northwestern.edu/netlogo/. Center for
configuration, Int. J. Electr. Power Energy Syst. 71 (2015) 391–402. Connected Learning and Computer-Based Modeling, Northwestern University,
[20] Y. Wang, G. Wei, H. Yang, H. Chen, Z. Ouyang, D.J. Hill, Novel protection scheme of Evanston, IL.
single-phase earth fault for radial distribution systems with distributed generators, [32] R.D. Zimmerman, Hsiao-Dong Chiang, Fast decoupled power flow for unbalanced
IEEE Trans. Power Deliv. 99 (2016) 1–1. radial distribution systems, IEEE Trans. Power Syst. 10 (November (4)) (1995)
[21] A.N. Albagli, D.M. Falcão, J.F. de Rezende, Smart grid framework co-simulation 2045–2052.
using HLA architecture, Electr. Power Syst. Res. 130 (2016) 22–33. [33] M.B. Biggs, J.A. Papin, Novel multiscale modeling tool applied to Pseudomonas
[22] N.G. Tarhunia, N.I. Elkalashyb, T.A. Kawadyb, M. Lehtonenc, Autonomous control aeruginosa biofilm formation, PLoS One 8 (October (10)) (2013).
strategy for fault management in distribution networks, Electr. Power Syst. Res. 121 [34] www.maisy.com/.

239