ARTIFICIAL NEURAL NETWORK BASED POWER SYSTEM STABILIZER ON A SINGLE MACHINE INFINITE BUS SYSTEM MODELLED IN DIGSILENT POWERFACTORY AND MATLAB

Electrical Engineering: An International Journal (EEIJ), Vol. 2, No. 2/3/4, December 2015
DOI : 10.5121/eeij.2015.2401 1
ARTIFICIAL NEURAL NETWORK BASED POWER
SYSTEM STABILIZER ON A SINGLE MACHINE INFINITE
BUS SYSTEM MODELLED IN DIGSILENT
POWERFACTORY AND MATLAB
Ali Kharrazi
School of Engineering and Information Technology
Murdoch University, Western Australia
ABSTRACT
In this paper the use of artificial neural network in power system stability is studied. A predictive
controller based on two neural networks is designed and tested on a single machine infinite bus system
which is used to replace conventional power system stabilizers. They have been used for decades in
power system to dampen small amplitude low frequency oscillation in power systems. The increases in
size and complexity of power systems have cast a shadow on efficiency of conventional method. New
control strategies have been proposed in many researches. Artificial Neural Networks have been studied
in many publications but lack of assurance of their functionality has hindered the practical usage of them
in utilities. The proposed control structure is modelled using a novel data exchange established between
MATLAB and DIgSILENT power factory. The result of simulation proves the efficiency of the proposed
structure.
KEYWORDS
MATLAB, DIgSILENT PowerFactory, Power System Stabilizer, Artificial Neural Network
1. Introduction
Small amplitude low frequency oscillations have been observed in power systems since 1960’s
[3]. They are the result of kinetic energy exchange between generation units. These oscillations
have been known as electromagnetic mode of oscillation and if they do not dampen properly,
they could affect the power quality or even cause the isolation of a part of power system. They
may happen in two modes, local and inter area. Local mode is when one generator swings
against the rest of the power system and inter area is when a group of generators in one area
swing against a group of generator in another area. From control structure point of view, the
phase lag exerted to the system as a result of the controller of the excitation system will
originate these oscillations [12].
In order to dampen these oscillations a component in phase with speed deviation should be
added to the mechanical torque. This has been achieved by use of conventional lead lag
controllers known as Conventional Power System Stabilizer (CPSS). CPSS have been used in
many utilities however their lack of flexibility to changes of power system has questioned their
efficiency. Many new control strategies have been proposed to achieve more adaptive
stabilizers.
Artificial intelligent techniques have been the focus of many researches [13]. In this paper a
predictive controller based on two artificial neural networks is proposed. The proposed model is
simulated on a novel data exchange platform using MATLAB and POWERFACOTRY data
exchange.
Simulation software packages have been used by engineers in order to investigate the behaviour
of systems and validate the functionality of new designs without running detrimental

2
experiments on real systems. In Electrical powers engineering various useful software packages
have been deployed by engineers, operators and students to study the behaviour of power
systems in various operating conditions. Along with the advances in equipments and
technologies in power systems, more complicated and enhanced simulation software tools with
more capabilities and applications have been introduced. The interface between two simulations
software where each software package is used for specific area makes the optimum use of them.
The Dynamic Data Exchange supported by operating systems creates and environment where
the real time data exchange between different software could be achieved [1].
In this paper a simulation platform using DIgSILENT PowerFactory and MATLAB Artificial
Neural Network Toolbox is proposed. PowerFactory is a powerful simulation tool for power
systems. Many conventional control loops could be modelled in PowerFactory but it does not
support advanced and intelligent control strategies. On the other hand it is possible to link
control blocks to MATLAB scripts with in control loop in PowerFactory [2]. In each time step
during simulation a common data base between MATLAB and PowerFactory is shaped where
PowerFactory exports desired data to MATLAB and calls the scripts which are saved as m file
in MATLAB environment. When the codes have been executed PowerFactory imports the result
of the calculations.
The power system under study is a Single Machine Infinite Bus (SMIB) modelled in
PowerFactory which is linked to Artificial Neural Network based Power System Stabilizer
coded in MATLAB using Neural Network Toolbox.
In next section of the report the application of Power System stabilizer (PSS) and different
control strategies used as PSS will be introduced. Next section describes the Power system
model in PowerFactory. In section 4 the data interface between the two software packages is
explained. Section 5 explains the Artificial Neural Network based Power System Stabilizer
(ANNPSS) created in MATLAB. Next section presents some simulation and results followed by
conclusion.
2. Power system stabilizer
Power system stabilizers have been used for decades in power systems to enhance the transient
stability of power system [3]. They are employed in order to dampen the small amplitude low
frequency oscillations observed in power systems. These oscillations in rotor speed and rotor
angle of generation units are result of phase lag imposed by excitation control system,
generation and power system.. The frequency of these oscillations is in range of 0.5 to 3 Hz.
They may persist in the power system and if does not dampen properly they may result in poor
power quality or even the isolation of generation unit or a part of power system.
In order to add enough dampening to the system, a component in phase with speed deviation is
added to electrical torque in rotating mass. The conventional power system stabilizers are
embedded through Automatic Voltage Regulator and the input is speed or speed deviation of the
rotor. The output is stabilizer signal (UPSS). In Figure1 the block diagram of a conventional
power system stabilizer (CPSS) is shown.
Figure 1 Conventional Power System Stabilizer

3
As seen in Fig.1 the CPSS consists of a gain block, a washout filter to filter out the steady state
voltage and two lead lag blocks to compensate the phase lag of the system. The parameters of
lead lags block should be tuned properly to achieve desired responses.
Although CPSS has been used widely in many utilities with good outcomes, they suffer from
drawbacks of using conventional controllers such as dealing with uncertainties. On the other the
hand complexity of power systems and its nonlinear nature as well as constantly changing
parameters of the system has motivated researches to propose more advance control strategies.
Many adaptive and robust control structures have been proposed to be used as Power System
Stabilizer [4] [5]. In many publications Artificial Intelligent control techniques have been used
to enhance the stability of power systems including tuning of the parameters of CPSS [6] or
replacing the conventional controller with Artificial Neural Network (ANN) structures [7] [8].
The flexibility and adaptablity of ANNs due to large number of parametrs and training
capability, have propmted researches to introduced Artificial Neural Network Based Power
System Stabilizers.
3. Power System under Study
The power system in this paper is a Single Machine Infinite Bus which is modelled in
PowerFactory shown in Figure 2. This topology is used to study the local mode of oscillation. It
is a single generator unit connected to an infinite bus through a transmission line. The infinite
bus is the equivalent of the rest of power system in form of a bus with fix voltage.
Figure 2 Single Machine Infinite
The control block of synchronous generator is modelled as composite model in PowerFactory.
Different control structure such as governor, AVR and PSS are embedded in composite model.
This structure is shown in Figure 3.
Figure 3 Composite model Signal Interconnection

4
Each slot in the composite model is linked to a control block known as common block. Now the
PSS block is linked to a CPSS shown in Fig. 1. This composite model could be changed and
modified by user. New blocks could be added and the signal interconnection could be changed.
4. PowerFactory MATLAB Interface
Classic control structure could be modelled in PowerFactory. These models could be embedded
in composite model shown in Fig. 3. But more complicated control structure could not be
simulated in PowerFactory. These constrains in simulation software have prompted the interface
of different simulation tools where each tool has a specific application. In [1] the process and
requirments of interfacing transinet simulation software which use mathematical algorithm to
simulate electromagnetic phenomena are illustrated.
In this paper a Neural Network based controller is replaced with the CPSS in Figure 3. A
common block is added to the composite model shown in Figure 3, which is linked to a
MATLAB script. In [2] the process to link a common model in PowerFactory to MATLAB
script is illustrated. The new signal interconnection with MATLAB interface block is shown in
Figure 4.
Figure 4 New Block for MATLAB Interface
Since MATLAB codes are to replace PSS in control loop, signals connected to this new block
are speed from Synchronous Generator and UPSS to the AVR block. In order to test the proper
functionality of the data interface the block diagram shown in Figure 1 was modelled in
MATLAB Simulink and the result of simulation were compared with the block modelled in
PowerFactory. The block modelled in Simulink is shown in figure 5.
Figure 5 CPSS modelled in SIMULINK
The validation of correct data exchange is done through a simulation of three phase fault in the
transmission line. The result of two CPSS block from MATLAB and POWEFACTORY are
shown in figure 6.

5
Figure 6 Result of Simulation Comparing Simulink and PowerFactory Blocks
It could be seen in Figure 6 that the simulation result done through data exchange between
POWERFACTORY and yield identical results to the simulation done by blocks in
POWERFACTORY.
5. Artificial Neural Network based Power System Stabilizer
(ANNPSS)
Although conventional and classic control structure could improve the transient stability of
power systems, but lack of adaptivity and inherent drawbacks in the design such as uncertainty
has cast a shadow on their performance. In addition the increasing complexity and demand of
power system has pushed them to operate at stability limits. Thus new intelligent and adaptive
controllers are inevitable.
Figure 7 ANNPSS Control Structure

6
Controllers based in ANN have been proposed in many publications. In this paper the control
strategy proposed in [9] is adapted to enhance the stability of power system. The proposed
structure is a predictive controller which use two neural networks in order to achieve control
goal in nonlinear systems: Neural Controller and Neural Identifier. The proposed structure is to
design an optimal controller to minimise a cost fucntion. Same topology has been used in [10]
to realize a power system stabilzier. The novelty in this paper is the software tools employed
which eases the training and simulation of the proposed system. The control structure of
ANNPSS is shown in Figure 5.
The Plant in Figure 7 is the power system as seen from point of view of PSS. The input to the
plant is the stabilizer signal and the output is the speed. This control structure is modelled in
MATLAB and is linked to PowerFactory. In composite model in PowerFactory two interfaces
are added for each Neural Network as shown in Figure 8. The two neural networks in the
control loop are illustrated in the following sections.
Figure 8 Neural Controller and Identifier Interface between MAT|LAB and PowerFactory
5.1 Neural Identifier
The main aim of Neural Identifier (NI) is to approximate the behaviour of the system. In order
to achieve this goal Non-linear Auto-Regressive Moving Average (NARMA) [11] model which
is a representation of nonlinear sytems is used. Based on NARMA model the output of a
discrete system is a function of the current and previous values of system inputs and outputs as
shown in Equation 1.
The NARMA model could be realized by feed-forward Neural Network which is modelled
using MATLAB Neural Network Toolbox (nntool). A Nonlinear Autoregressive with External
Input network (narxnet) including one hidden layer with ten neurons and output layer with one
neuron is created in MATLAB. This network could be seen in Figure11.
The task of this network is to learn to predict the output of the plant using the current and
previous values of input and output which are stabilizer signal and speed of rotor. These signals
are fed to the Neural Identifier Block in composite model shown in Figure6. The two following
sections illustrate the training and validating the identifier.

7
Figure 9 Neural Identifier Network created by nntool
5.1.1 Training Neural Identifier
The NI should be trained with adequate learning data in order to be able to predict the output of
the plant. The learning data which include stabilizer signal and speed should be chosen
adequately to give enough information for proper learning. The signals used, should excite all
oscillatory modes of the system. In order to achieve this MATLAB’s idinput command which
generates signal for identification is used. The signal shown in Figure 12 is created using the
mentioned command. This signal is sum of sinusoid signals with different frequencies in order
to excite all the modes of the plant.
Figure 10 Signal Generated for Identifier Training
This signal is send to PowerFactory using the MATLAB Interface block shown in Figure 10 as
UPSS signal during simulation. After running the simulation the speed of generator signal is
saved as CSV file in PowerFactory. This CSV file is loaded in MATLAB and is used as the
target data for offline training of Neural Identifier along with the signal shown in Figure 10. The
training process is done using train command in MATLAB. Online training could be performed
as well. In order to perform online training the Neural Identifier network should be embedded
into the MATLAB codes linked to Neural Identifier Block in Figure 8. The input signals to the
Neural Identifier Blocks are the signal coming from MATLAB Interface Block and speed signal
coming from Synchronous Generator Block.
The cost function of training is shown in Equation 2.
In Equation 2 w(k) is the output signal (rotor speed) of the real plant in modelled in
PowerFactory and is the output of Neural Identifier.
5.1.2 Validating Neural Identifier
In order to validate proper training of Neural Identifier, a simulation is run to compare the
output of the NI in MATLAB and the plant in PowerFactory. In the simulation a three phase
fault in transmission line in 0.2s with duration of 500 ms and a 10 % for increase in load in 8s
are scheduled. The result of simulation is shown in Figure 11. As seen in Figure 11 the output
of Neural Identifier is similar to the speed signal in PowerFactory.

8
Figure 11 Comparison of Neural Identifier Output and PowerFactory Model
5.2 Neural Controller
The Neural Controller (NC) should be trained to create the optimum stabilizer signal. A time
delay Neural Network created using nntool in MATLAB, is used as Neural Controller shown in
Figure 12.
Figure 12 Neural Controller
This network is embedded in MATLAB codes which are linked to the Neural Controller Block
in Figure 8. The training of NC is done using the NI output. The NI could predict one step ahead
of future value of the system response using the previous stabilizer signal and current and
previous speed signal. The predicted value is compared with desired value which in this case is
1 p.u since we want the speed to be fix at all times. So the cost function for training of NC could
be defined as shown in Equation 3.
In Equation 3 the cost function is the mean square error of desired value compared to NI
prediction. Since the cost function in Eq.3 is not a function of NC output, the training algorithm
in MATLAB nntool could not be used which use the derivative of cost function w.r.t. network
output. In order to train the NC the output mean square error of NI is defined and this error is
propagated back through NI using chain rule. The output error and back propagated error are
shown in figure 13.
Figure 13 Error of Identifier Output with Desired Value and Back Propagated Error

9
This process is embedded in MATLAB codes linked to Neural Identifier Block. In each time
step the signal is fed to the system as UPSS signal. NI predicts the next time step of output of
the system and the mean square error of predicted value and desired value is calculated. This
error is back propagated through NI and the error of stabilizer signal which is the output of NC
is obtained. This Signal is used to train the NC.
6. Simulation and Results
In order to validate and test the functionality of the proposed control structure some simulations
are performed. After training of NC the Neural Identifier Block is disabled and the output of
Neural Controller block which is linked to MATLAB is connected to AVR block as UPSS
signal. All the simulations are run in PowerFactory with active data interface with MATLAB.
In the first simulation shown in Figure 14, a three phase fault lasting for 500 ms at 0.2s and 10
% increase in load in 8s are scheduled.
Figure 14 ANNPSS and CPSS Comparison to 500 ms Three Phase Fault and 10 % load increase
In Figure 15 12 the result of simulation of three phase fault which lasts for 1s is shown.
Figure 15 ANNPSS and CPSS Comparison to 1s Three Phase Fault
The last simulation is performed while the local load of the power system is in maximum. The
results are shown in Figure 16.
Figure 16 ANNPSS and CPSS Comparison to 500 ms Three Phase Fault in Full Loading

10
In all the simulation the response of rotor speed to CPSS modelled in PowerFactory is
compared with response to ANNPSS modelled in MATLAB. It could be seen from the
figures that ANNPSS has better response and could regain steady state in a shorter time.
7. Conclusion
In this paper a new platform for data interface between MATLAB and DIgSILENT
PowerFactory was employed to run the simulation of an Artificial Neural Networked based
Power System Stabilizer. The interface between the two software packages makes optimum use
of each one in their specific area of applications. The proposed Neural Network control strategy
is modelled and trained in MATLAB Environment using the data shared with PowerFactory
during simulation. The result of simulation validates the effectiveness of proposed control
structure while the data interface facilitates the process of simulation.
REFERENCES
[1] S. Filizadeh, M. Heidari, A. Mehrizi-Sani, J. Jatskevich and J. Martinez, "Techniques for
Interfacing Electromagnetic Transient Simulation Programs With General Mathematical
Tools IEEE Taskforce on Interfacing Techniques for Simulation Tools," Power Delivery,
IEEE Transactions on , vol. 23, no. 4, pp. 2610 - 2622, 2008.
[2] M. M. A. F. L. G. T. Shadi Khaleghi Kerahroudi, "Application and Requirement of
DIgSILENT PowerFactory to MATLAB/Simulink Interface," in PowerFactory
Applications for Power System Analysis, Springer, 2014, pp. 297- 322.
[3] E. Larsen and D. Swann, "Applying Power System Stabilizer, Part1; General
Concept,Part2; Performance Objective and Tuning Concepts, Part3; Practical
Consideration," IEEE Transactions on Power Apparatus and Systems,, Vols. PAS -100, no.
6, pp. 3017-3046, 1981.
[4] Q. Zhao and J. Jiang, "Robust controller design for generator excitation systems," Energy
Conversion, IEEE Transactions on , vol. 10, no. 2, pp. 201 - 209, 1995.
[5] B. Wu and O. Malik, "Multivariable Adaptive Control of Synchronous Machines in a
Multimachine Power System," Power Systems, IEEE Transactions on (Volume:21 , Issue:
4 ), vol. 21, no. 4, pp. 1772 - 1781, 2006.
[6] Y.-Y. Hsu and C.-R. Chen, "Tuning of power system stabilizers using an artificial neural
network," Energy Conversion, IEEE Transactions on (Volume:6 , Issue: 4 ), pp. 612 - 619,
Dec 1991.
[7] A. Sharaf, T. Lie and H. Gooi, "Neural Network Based Power System Stabilizers".
[8] Y.-M. Park, M.-S. Choi and K. Lee, "A neural network-based power system stabilizer
using power flow characteristics," Energy Conversion, IEEE Transactions on , vol. 11, no.
2, pp. 435 - 441, 1996.
[9] Y.-M. Park, M.-S. Choi and K. Y. Lee, "An Optimal Tracking Neuro-Controllerfor
Nonlinear Dynamic Systems," IEEE TRANSACTIONS ON NEURAL NETWORKS, ,
vol. 7, no. 5, pp. 1099-110, 1996.
[10] W. Liu, G. K. Venayagamoorthy and D. C. Wunsch, "Design of an adaptive neural network
based power system stabilizer," Neural Networks, Elsevier, vol. 16, no. 5, pp. 891-898,
2003.
[11] S. Chena and S. A. Billingsa, "Representations of non-linear systems: the NARMAX
model," International Journal of Control, vol. 49, no. 3, pp. 1013-1032, 2008.

11
[12] F. Demello and C. Concordia, “Concepts of Synchronous Machine Stability as
Affected by Excitation Control,” Power Apparatus and Systems, IEEE Transactions
on , Vols. PAS-88, no. 4, pp. 316 - 329, 1969.
[13 ] L. Hassan, M. Moghavvemi and H. Mohamed, “Power system stabilization based
on artificial intelligent techniques; A review,” in Technical Postgraduates
(TECHPOS), International Conference for, Kuala Lumpur, 2009.

ARTIFICIAL NEURAL NETWORK BASED POWER SYSTEM STABILIZER ON A SINGLE MACHINE INFINITE BUS SYSTEM MODELLED IN DIGSILENT POWERFACTORY AND MATLAB

More Related Content

Similar to ARTIFICIAL NEURAL NETWORK BASED POWER SYSTEM STABILIZER ON A SINGLE MACHINE INFINITE BUS SYSTEM MODELLED IN DIGSILENT POWERFACTORY AND MATLAB (20)

More from hugoshan513 (20)

Recently uploaded (20)

ARTIFICIAL NEURAL NETWORK BASED POWER SYSTEM STABILIZER ON A SINGLE MACHINE INFINITE BUS SYSTEM MODELLED IN DIGSILENT POWERFACTORY AND MATLAB