International conference on Signal Processing, Communication, Power and Embedded System (SCOPES)-2016

LOAD – SOURCE ANALYSIS AND

SCHEDULING IN HYBRID SMART
GRID USING MULTILAYER ADAPTIVE
GNN ALGORITHM
Sarin CR M.E, M.B.A, (Ph.D) Syam PS Nair
Assistant Professor Student, MD in Electrical Engineering
Vimal Jyothi Engineering College, Kerala, India University of Southern Queensland - Australia
[email protected]
Dr. Geetha Mani M.E., Ph.D.,
Santhosh Krishnan M.E Associate Professor,
Assistant Professor, School of Electrical Engineering, VIT University,
Madawalabu University,Bale, Ethiopia India
[email protected] [email protected]

through large scale hydropower plants or thermal

power plants or nuclear power plants, small scale
Abstract - The solar and wind power has renewable energy sources have been added to the
recently become a potential option in power grids in online mode which generates energy for
systems and act significantly to meet the user and feeds back the excess energy to the
thepenetration of demands. The present growth grid. Most of the systems are distributed in nature
of such renewable energy sources has shown within various sessions of grid, irregular
anexponential increase in on-grid generation geographical areas which may be at a point far
systems. The high penetration of such systems distance from the centralized controller. The
help a grid to effectively meet its load popular small range renewable energy sources like
requirements during an irregular demand but solar and wind has low power and is in single
also create some disturbances in grid due to phase still provide an effective method to meet the
frequent addition and detachments of load demands with less transmission losses as the
loads/sources. The renewable energy sources generating points will be nearer to the load sites.
that usually works in on-grid mod is to be Almost every system has an irregular source of
attached and cut down from the grids without generation, and the magnitude of the energy
creating disturbances in stable grid. Another generated vary from source to source and time to
important requirement is the effective load time.
management with less transmission losses. The Whenever such systems are attached or
proposed system introduces a method for detached, sudden changes create some disturbances
optimized source additions and effective load in the system. This disturbance may affect the
scheduling without disturbing the stability of the effective stable working of the grid. So, as to
system. It uses a three-layer metaheuristics achieve a stable characteristic within the grid, the
multidimensional algorithm, composed of proposed system facilitates distributed sensors to
Adaptive Hopfield network which is used to achieve present status variation and uses and
identify the gridand a Genetic Algorithm to intelligent adaptive methods analyses the variables
identify the optimist load scheduling. and to provide optimum solution. There are
numerous dynamic variables in a power system to
Keywords: Artificial Intelligence, ANN, GA, Power be analyzed. The proposed system analyses the
system, Dynamics, Stability, Soft Computing major transient dynamic parameters like base load
changes, changes in power factor, source additions
I. INTRODUCTION and detachments, variation in primary sources,
Use of renewable energy sources have fault strike voltages and overloads. It also
been increased rapidly in recent years due to the determines the transmission losses and provide the
changes in availability of fossil fuel, high cost of best cost estimation for the management of sources.
the fuel and the increasing environmental concerns. These parameters have been measured for
This has created a platform for grid based system a definite time period and the variations of each
which has hybrid source generators. Though the parameter is monitored. Once the parameters have
major share of the power generation is done been monitored, the rate of change of variations

978-1-5090-4620-1/16/$31.00 ©2016 IEEE 1900

 International conference on Signal Processing, Communication, Power and Embedded System (SCOPES)-2016

have been identified. To achieve the better stability, variations and Suites variations using particle
the parameters have been analyzed in detail and Swarm Optimization in a wind power solar
they have been sorted on the basis of rate of integrated power system. Pei the peak Wu, Student
change of variations. A threshold function is time, the Wang, and Sheng Bi has analyzed
implemented to identify the dangerous or available opportunities for the implementation of
considerable variations. The process of variable soft computing methods in load implemented. It
analysis is determined using neural network. The forecasting. Abderrezak Laouafi, Mourad
advantages of Hopfield network are the variation of Mordjaoui and Djalel Dib yes derive system using
east point of the grid is reflected at a point. Once artificial neural network for one hour effect electric
the dynamic behavior of the system is analyzed load estimation in wind solar hybrid power
using the Hopfield networks, the corrective generating system attached to a grid system.
measures have been implemented. III. Dynamic parameter evaluation
The Hop-field function has been set with a Most of the present models of renewable
matrix whose values have been are presented with energy sources such as wind or solar energy works
two states 1 and -1 so that its variable may be only for a fraction of actual requirements and an
assigned with one of these values based on the another primary source is added for reliable and
nature of change of variable. The Hopfield matrix uninterruptible supply of power. The primary
is found at each user node and set with the weights source will be of large power rating, usually
based on the nature of the faults. Here the whole generate energy from hydropower, thermal power
network has been modified so that instead of using or large scale wind farms. Considering the cases of
single weight to all the parameters, different small scale renewable energy sources such as wind
parameters are assigned with different weight or solar, the capacity of such systems falls under 5
fractions. This type of graded neuron model may to 20 Kilowatts, specially designed to meet the
have different weight fractions based on the nature supply within domestic requirements. The average
of variable. Still the cost effective optimization is reliable time period of solar system is 10 hours
also a necessary criterion for the design and with 5 Kilowatts/hr. and has of the wind is 8 to 16
operation of a power system. A knack pack based hours with an average power generation of 4
problem solving is used for selective load Kilowatts/hr. The average conception from a
scheduling and source scheduling with least value domestic user may vary from 3 Kilowatt to 10
of primary source. The knack pack availability is Kilowatt at the peak time.
solved using genetic algorithm which set up an Since during the peak time, system make
optimum schedule. There added an adaptive weight use of primary source as the source of energy and
control within Hopfield network which provide real in all other conditions, its feeds energy from the
algorithms or weight adjustment so that the system renewable energy sources, and also feeds the
control the dynamic behaviors based on both excess of energy to grid. The user behavior may
present parameters status and also the cost effective vary from person to person and time to time so that
solution from genetic algorithm. the person may feed energy in either side from time
II. Literature review to time the respective of the area or the status of the
Tridib Kumar Das, Ganesh Kumar primary source availability. As such, it is important
Venayagamoorthy and Usman O. Aliyu shown that to monitor search changes in the grid, forecast the
the inter area oscillations are critical to optimal energy requirements, avail the possible source
power flow and stability in a process system and options and schedule the loads based on the
also for the better Optimisation, it is recommended availability of power. eliminated among the first
to use multi layer algorithms. H. R. Cai, C. Y. system make use of a SCADA based monitoring
Chung, and K. P. for the of nonlinear that the system and controller which frequently monitor the
consideration Moursi, and constraint in optimal dynamic parameters in the grid, unless them using
power flow problems is much important and artificial intelligence and control the source and
identifying the Global optimum solutions of such load options. For the analysis, each user is taken as
nonlinear variables can be obtained by evolutionary a node and analysis is performed identify the nature
Zeineldin provided multidimensional algorithms. S of the node. Since the behavior of each user is
Lee has developed a system solar, theof aMember, reflected in the grid to other nodes, it could be
XiaojunZhang, Xiaoyu possible options for using taken as each node is interconnected. For such
genetic algorithm in classification of nonlinear analysis, Hopfield network is the best option.
independent variables. Surour Alaraifi, M. S. El One of the major disadvantage of the
Moursi, and H. H. Zeineldin provided possible conventional Hopfield analysis is that each node
implementation of multilayer algorithm in control can have of maximum of two stage values, one
of an induction motor using mixed integrated non complement to another. node may have a number
linear programming and particle Swarm of different parameters, hence a modified version
Optimisation. Duy C. Huynh and Loc D. Ho has of Hopfield network called as gradient Hopfield
derived solution for the optimal control of load network is implemented. A number of parameters

1901
 International conference on Signal Processing, Communication, Power and Embedded System (SCOPES)-2016

have been measured which includes the real power nature. To overcome these difficulties, the Hopfield
and reactive power consumption, the availability of network have been modified and alterations have
additional renewable energy sources, the amount been made with respect to the requirements. First,
power tapped from the form the renewable energy the gradient based Hopfield network has been
sources, amount of power supply to the grid and selected and a multiple row matrix have been
the available primary power. selected for weight adjustments.
All these parameters have been measured
at each node and these variables have been sent
along with the node ID central SCADA processing
system. The proposed system implements as
simulated model of 75 users and the data
consumption of 700 samples have been taken.
Once Decembers have analyzed, the minimum,
medium, and peak load behavior of the grid system
is identified using the artificial intelligence
algorithm and the source additions have been
controlled, implemented load scheduling if
necessary. It also determines the transient behavior
of the system whenever source/load is attached or Figure I
detached to the system, by analyzing the power A. Hopfield Network Layer
rating at two consecutive time intervals. If the node The power consumption at an individual
to is beyond the threshold limits, respected nodes node is set as the primary variable for Hopfield
cut off from the system and add it again. One of the matrix. The variations of the power are made
major advantage of the proposed system is that, it positive if it is consumed which represents the node
analyses the complete nodes actions together to as a load point and it is made negative if the node
determine the transient behavior instead of act as a source to the grid. For the modelling, the
analyzing individual load sections. The cost maximum amount of power consumption and
function analysis is implemented as a feedback generation is said to be 10 Kilowatts. The variation
adaptive weight to the Hopfield network which of power is expressed in per unit using a boundary
helps in intelligent load/ source scheduling. matrix whose values may change from -1 to +1 as
IV. Adaptive GNN Algorithm gradients. The reactive power construction is also
The algorithm used in proposed system is represented in per unit using and another Bond
a three-layer algorithm which is a combination of Matrix whose value may change from -1 to +1 as
genetic algorithm and an artificial neural network. gradients. The positive value of the reactive
The first part of the algorithm, which is basically a represents a leading power factor and the negative
search algorithm, is used for determining the values represents lagging power factor. The real
transient behavior or dynamic behavior on each power gradient is also measured in a function same
node. differentiate the behaviors that exceeds the as that of a apparent power calculation.
threshold limits. It also determines the correlation
between each variable to N other variables. The
second part has a selection algorithm which
determines node to be controlled. The third part of
the algorithm implements a cost function based
adaptive feedback to the respective same manner
meet the requirements. If the node functions are
beyond the control, scheduling is recommended.
Hopfield neural network algorithm has
been selected as the base algorithm in layer 1. One
of the major advantages of Hopfield network is that
each node of the system is interconnected to all
other nodes so that any variation in one node may Figure II
affect the other nodes. Similarly, in a power The primary power value is said to be the
system, the variation in 1 node will always affects fourth variable in Hopfield matrix, the gradient
the other counts. Still, one of the major drawback may vary from -1 to + 1 in such a way that
with their Hopfield at work is that it only allows 2 whenever the primary energy is the base source, the
States of processor. It also creates some difficulty value of the grid in matrix is minus one and if the
as a common weight is applied positive. The system is independent of additional sources of
source / load addition and attachment on each node power, the value of the grid matrix is 1. Weight
is selected, and it may be inappropriate to apply the matrix will change from time to time the per the
same weight to all the values irrespective of their adaptive variations from the final iteration which is

1902
 International conference oon Signal Processing, Communication, Power and Embeddedd System (SCOPES)-2016

found using a genetic algorithm.. The selection In an electrical network m most of the nodes are
description of the matrix array is m made in such a interconnected and a variation in one node may
way that the best array provides higher leading rtant select nodes which
reflect in another, it is import
power factor, feed maximum enerrgy back to the have direct impact on the diisturbance. This cannot
main grid and primary energy consuumption is at its be identified using a field neetwork still, identifying
minimum. Initial weight Matrix havve been set as [ the behavior of each node to another using the
1111]. same Hopfield network is time consuming.
Since its not has its ownn behavior, the Overcome this, different froom conventional neural
variations have been denoted usingg an undirected network algorithm where is single solution is
graph node. determined using iterative thhreshold function, here
1. Define the number of userr nodes N where the threshold function is vvaried form layers to
represent user. layers, iterative set is madde and all number of
2. The proposed Hopfield nnetwork consists nodes that satisfies the threshold criteria is taken.
of 75 * 75 =5625 neuronss and it's neuron Once the nodes have been identified and
may take average value frrom - 1 to + 1. correlation between them is found, and it is given
The weight of matrix is sett as as a matrix to the genetic algorithm identify the
first node which is to be conntrolled immediately or
in sequence.
9. Determine the poopulation of selected
nodes which he haas been decided using
Hopfield network.
i= 1 to Ti l implemeented. What of 10. Derive the fitness function, reject all
variable such as real ppower, reactive nodes that does not meet the fitness
power, apparent power, primary power function
consumption. Where AW is the adaptive 11. Start crossover w with the most desired
weight obtained from thee final iterative solution.
analysis. The address is taaken only from 12. Apply mutation to each chromosome and
node to node, hence the vvalue of i and j find the best possiblle option.
cannot be the same The criteria for setting up thee fitness function of the
3. Initialize the network so thhat the value of chromosome is set based on two parameters,
i,j, l, simulated function arre all set to zero. 1. Source load requirements
2. Cost beneffit function
4. Start the iteration and fiind the possible If Both functions are beyond the threshold
values of simulated functioon value, cut off the respective nodes. If it is nearer to
the threshold value, a feedbaack is send to adaptive
network which adjust the w weight of the respective
node with best possible opptions. If the option is
negative even in the secondd iteration, the node is
5. Fitness function is derivved using with cut off. A model of the propposed system has been
respect to the variations in the network developed, it is simulatedd using Matlab and
using possible variations are analyzed. For the
F = Ti simulation, readings from 775 different uses have
been considered where eaach user may have a
6. Create second layer of neuurons that means domestic uses maximum 100 Kilowatts. The total
the fitness function thrreshold criteria capacity of the grit is set to 1 megawatt out of
where the number of layerrs in the second which the 365 megawatt haas been satisfied using
layer is limited to the halff of the previous renewable energy sources w which has been attached
layer. 75 different nodes with a caapacity of 5 Kilowatt at
7. ma and Maxima
Identify the Global minim each node. If the constructioon from average user is
and also plot local minim
ma and Maxima beyond Kilowatt, the user maakes use of the primary
for it's cluster. supply for additional poweer. If it is less than 5
Kilowatt, the user needs arre met with renewable
One of the major drawback off Hopfield based
energy and additional energgy tap forward to the
neural network is that the time takenn for iteration is
grid. To get the readings ffrom each node, smart
too large as there is an N*N matrixx. To overcome
meter is attached at each nodde.
this difficulty, instead of using a single optimum
V. Results and Disscussions
function identification , multi neurron selection is
The nature of uuser consumption is
implemented which has very lless correlation
identified at various time inttervals and is plotted in
coefficient . It is not necessary too evaluate every
graph 1.
node but the nodes with large dynaamic variations.

1903
 International conference oon Signal Processing, Communication, Power and Embeddedd System (SCOPES)-2016

Thedynamic behavior of a uuser for the 700 times

samples is also measured and is analyzed the
dynamic behavior of each user on is described
using the correlation with tthe very next value. It
could be shown that in som me cases the variation
may reach up to 80 percentaage increase/ decrease.
This behavior can be taken as a transient behavior
and user one shows such behhavior at least 10 times
for this 700 samples. It is shoown in graph 4.

Grraph 4
Graph I
And another set of you shoould behavior is
also shown in graph 2.

Graph 5
Graph 2 With respect this changes, the dynamic
The dynamic behavior off each user with behavior of each user with rrespect to or other uses
respect and another user is meassured using the is also analyzed and found iff this variation does not
correlation. This pattern describes thhe variation that go beyond a definite thrreshold value. The
may occur in each node due to anoother node. The simultaneous correlation bettween 3 different users
positive correlation explains that, aas the voltage in is found in graph 5.There aare two constraints, the
the first node increases, the voltage in the other not time of occurrence of dynamic behaviorand
also increases and vice versa. Thhe correlation of nodewhere the event occurss. To find the optimal
user is plotted in graph 3. solution, after a number of iterations 9 sets of time
samples have been taken whhich shows the highest
range of dynamic variation and the user behavior
during this time samples hass been plotted in graph
6.

Grraph 6
Graph 3 The behavior of eacch user during this time
From the graph, it could be dooubted that the samples is represented in graaph 7. The address and
variation in voltage at one node duue to another is other values for the neural nnetwork is seen in class
very large, and is almost nonn independent. 7 and the variation usingg adaptive network is

1904
 International conference oon Signal Processing, Communication, Power and Embeddedd System (SCOPES)-2016

shown in graph 8. It could be noddeed that using the

adaptive feedback, the number off iterations and
error value is reduced.

Graph 110
The dynamic behavvior of a user which is
improved using adaptive G GNN algorithm is then
represented in graph 11. Thhe dynamic correlation
of user is also shown in ggraph 12. It could be
analyzed that, after the adaaptive monitoring and
control, the system resource less transient behavior
throughout the time scale.

Graph 7

Graph 111

Graph 8

Graaph 13
The control over dynamic behavior is
achieved by effective load shedding of user nodes.
The load scheduling of usee of nodes for definite
Graph 9 time samples has shown in graph
g 14.
The complete dynamic beehavior of each
user throughout the time samples have been also
analyzed which is shown in graph 10.

1905
 International conference oon Signal Processing, Communication, Power and Embeddedd System (SCOPES)-2016

[3]. A particle swarm optim mization for economic

dispatch with nonsmooth coost functions, Published
in: IEEE Transactions onn Power Systems (
Volume: 20, Issue: 1, Feb. 2005 ), Page(s): 34 - 42,
31 January 2005.
[4]. K. Y. Lee and M. A.. El-Sharkawi Modern
Heuristic Optimization Techniques with
Applications to Power Systeems 2002 IEEE Power
Engineering Society (02TP1660).
[5]. Optimal Allocation off HTS-FCL for Power
System Security and Sttability Enhancement,
Published in: IEEE Transactiions on Power Systems
( Volume: 28, Issue: 4, Nov. 2013 ), Page(s): 4701
Graph 15
- 4711 ,ISSN: 1558-0679.
Using the final wireless obtained, load
[61]. G. Didier J. Levequee and A. Rezzoug "A
forecasting is also determined usinng Drift method
novel approach to determine the optimal location of
like for next 50 samples which is shown in graph
SFCL in electric power grrid to improve power
16.
system stability" IEEE Traans. Power underSyst.
vol. 28 no. 2 pp. 978-984 Maay 2013.
[7]. Duy C. Huynh and M Matthew W. Dunnigan,
“Off-line and on-line parrameter estimation of
induction machines: advaanced particle swarm
optimization algorithms annd advanced recursive
least-squares algorithms”, Lambert Academic
Publishing House, Germany,, 2011.
[8]. Short-Term Load Foreecasting Based on Big
Data Technologies, Pei Zhhang, Member, CSEE,
Senior Member, IEEE, X Xiaoyu,Still, the. Csee
Journal Of Power And Energgy Systems, Vol. 1, No.
3, September 2015
[9]. One-hour ahead electric load and wind-solar
Graph 17 power generation forecastingg using artificial neural
network, Published in: Renewable Energy
VI. Conclusion Congress (IREC), 2015 6th International, Date of
Hybrids algorithm model for load source Conference: 24-26 March 2015, Date Added to
analysis and scheduling in a hybridd power system IEEE Xplore: 21 May 2015, Electronic ISBN: 978-
is determined. The dynamic parameeter behavior of 1-4799-7947-9
each and every user is monitored, analyzed using [9]. A. Laouafi M. Mordjaooui and D. Dib "One-
artificial neural network methods suuch as Hopfield Hour Ahead Electric Loaad Forecasting Using
network, selected the most requeested parameter Neuro-fuzzy System in a Parallel Approach"
using genetic algorithms and the fiinal values have Computational intelligencce applications in
been sent as a feedback to scheduule the load and modeling and control Studdies in Computational
source. The proposed system has shown smooth Intelligencevol. 575 pp. 95-1121 2015.
operation and it has been proved as an effective [9]. A. Laouafi M. Mordjaooui and D. Dib "Very
model for scheduling. The combinattion of Hopfield Short-Term Electricity Dem mand Forecasting using
neural network and genetic algoritthm is modified Adaptive Exponential Smootthing methods.
with respect to the cause and the im
mprovement have [10]. E. Gonzalez-Romera M.A. Jaramillo-Moran
been implemented. and D. Carmona-Fernandeez "Monthly Electric

 Energy Demand Forecastiing Based on Trend
Extraction" IEEE Transactioons on Power Systems,
vol. 21 no. 4 pp. 1946-1953 N November 2006.
[1]. Das, Triadic, and Ganesh Vennayagamoorthy. [11]. A. Anvari Moghaddam m and A.R. Seifi "Study
"Bio-inspired Algorithms for thhe Design of of forecasting renewable ennergies in smart grids
Multiple Optimal Power Systeem Stabilizers: using linear predictive filterss and neural networks"
SPPSO and BFA." Conference Reccord of the 2006 IET Renewable Power Geneeration vol. 5 no. 5 pp.
IEEE Industry Applications Conference Forty-First 470-480 November 2011.
IAS Annual Meeting (2006). [12]. H. W. Park, Y. Y. Lee,, J. R. Jiang, “Study on
[2]. Venayagamoorthy, Ganesh Kum mar. "Intelligent big data center traffic man nagement based on the
Sense-making for Smart Grid Stabillity." 2011 IEEE zeparation of large-scalee data stream,” in
Power and Energy Society Geeneral Meeting Proceedings of Seventh Intternational Conference
(2011).

1906
 International conference on Signal Processing, Communication, Power and Embedded System (SCOPES)-2016

on Innovative Mobil and Internet Services in

Ubiquitous Computing (IMIS), Taichung, China,
2013, pp. 591–594.

1907