Supervisory Power System Stability Control Using
Neuro-Fuzzy System and Particle Swarm
Optimization Algorithm
Abdulhafid Sallama
School of Engineering and Design,
Brunel University,
Abdulhafid.Sallama@brunel.ac.uk
Maysam Abbod School of Engineering
and Design,
Brunel University,
Maysam.Abbod@brunel.ac.uk
A
bstra
ct- This paper describes the design and implementation of
advanced Supervisory Power System Stability Controller
(SPSSC) using Neuro-fuzzy system, and MATLAB S-function
tool where the controller is taught from data generated by
simulating the system for the optimal control regime. The
controller is compared to a multi-band control system which is
utilized to stabilize the system for different operating conditions.
Simulation results show that the supervisory power system
stability controller has produced better control action in
stabilizing the system for conditions such as: normal, after
disturbance in the electrical national grid as a result of changing
of the plant capacity like renewable energy units, high load
reduction or in the worst case of fault in operating the system,
e.g. phase short circuit to ground. The new controller led to
making the settling time and overshoot after disturbances
proved to be lower which means that the system can reach to
stability in the shortest time and with minimum disruption. Such
behaviour will improve the quality of the provided power to the
power grid.
Index Terms—Supervisory control power system stability
neuro-fuzzy logic sequential particle swarm optimization
(SPSO).
I.
INTRODUCTION
Supervisory expert control concentrates on general
information about the process and the controller. The
decision-making in the supervisory control system is related
to situations involving major disturbances, technical faults,
inappropriate human actions, and a combination of such
events [1]. In such events the established control algorithm
does not apply, and the planning for proper actions by the
controller depends on knowledge about the functional
properties of the system. The planning for a new control
strategy will depend on information about the process
characteristics in a specific situation. Such a system should be
capable of performing the following tasks: monitoring the
performance of the controller and the process, detecting
possible system component failure or malfunctioning and
replacing the control algorithm to maintain stability, and
selecting the appropriate control algorithm best suited for a
particular situation. Such a system can be formed in a closed
loop to provide a conceptualized hierarchical system which
consists of a supervision level as the highest hierarchical
level, and the basic control level as the lowest. In general,
more tasks can be handled by the supervisory control
algorithm [2]e.g. start up and shut down procedures, process
optimization, fault diagnosis, response to malfunctioning
Gareth TaylorSchool of
Engineering and Design,
Brunel University,
Gareth.Taylor@brunel.ac.uk
beha
viour
, pattern recognition start and stop parameters estimation, and
alarm handling procedure.
To analysing and studying any system it is very important
step need it is the modelling, because it is related to process
characterization and design studies. Earlier, it has been
thought that a complicated mathematical approach could
model a system more accurately, but this still has problems
when non-linear, complicated and undefined system are
encountered. Conversely, the human mind can easily reach a
very good result when they deal with very complicated
system such as food preparation, playing football, dealing
with the machine as driving the car in the off-road and so on.
All of this and he did not gives any attention to the
mathematical models that describe the processes in their
brain, In order to be able control them, and he still perform
very well from human experience. In the past few decades,
the possibility of creating models which function more like
human thinking, through fuzzy set theory proposed by [3].
Later, several authors conducted research into fuzzy
modelling, which divided into six different methods [4] :
 Verbalization or linguistics through interaction with the
human operator or domain expert [5], [6] and [7].
 Logic analysis of the input and output data, [8] and [9].
 Fuzzy implication and reasoning algorithms to identify
fuzzy models [10].
 Identification and self-learning algorithms fuzzy
modelling of (multi-input / single-output) MISO systems
[11].
 Learning signals to create a rule-base [12].
 Self-organizing fuzzy modelling algorithm to model the
system via on-line input and output data [13].
However, In order to increase the efficiency of the fuzzy
controllers, and covers some of the problems such as nonminimum phase processes. In this paper, a knowledge-based
fuzzy modelling approach is presented in an attempt to model
non-linear systems in general, and to be applied in particular
to the power electric network which is facing a several
problems as the connection and losing of a large load from
the grid at peak times. A supervisory self-monitors and
decision fuzzy logic control (SSMDFLC) structure which
includes the first level of network monitoring to note any
irregular change during normal operation. In the second level,
after noting the change, the logical analyzed program by
special comparisons in the MATLAB program it will be done,
than through that it can be recognize the fault type. As such,
in the third level it will take the appropriate decision depends
on the fault type, which was trained the Supervisory control
system in it previously [14].
II.
SUPERVISORY CONTROL
Supervisory is tracking and focusing on specific information
about the process and controller. Such a system must be
monitored to detect the system and the controller performance
disturbances in the controlled system also the change in order
to maintain the basic requirements such as the stability and
select the appropriate value of scale of factor for the fuzzy
logic controller, as best suited to the specific situation. Such
as this system consist of a hierarchy construction by three
levels described in Figure 1. The highest level of supervisory
control will be do all the decision-making observe any
failures in the system and diagnose the new case after fault
occur, to find out the fault type, based on this an appropriate
decision it will be taken. The level of observer is an interface
between the different levels of the control system and the high
level. The lower level is working to adjust for level control. If
a simple closed circuit using any control device parameters in
any system, there are certain behaviours are acceptable and
others are not.
primary or adaptive control, In order to get acceptable
behaviour of the system as a result of those actions. The
supervisory level must also have an alarm- handing facility,
so that when a fault occurs and is detected firstly an alarms
issued.
III. CONTROL DESIGN
The controller was designed in several stages using the
Matlab Fuzzy Logic Toolbox [15]. Mostly with FLC design,
the first stage to choose the correct input signal. In this paper
the generator speed deviation (Δω), and its derivative (Δώ)
are two signals considered like two inputs for (FPSS)
controller as shown in Figure 2. These two signals are used as
rule-antecedent (IF-part) in the formation of rule base, and the
output of controller (Δu) is used to represent the contents of
the rule consequent (THEN-part) in performing of rule base
[16], which is injected into the input of the excitation circuit
controller.
1
error
ke1
In1
Gain
du/dt
Decision- Making
Rule Base
Δu
c_error
Derivative
ko1
1
kc1
Gain1
Fuzzy Logic
Controller
Out1
Gain2
Figure 2. Block diagram for FLC controller.
Observer
+
FLC Controllers
Power system
(process)
_
Figurer 1 Supervisory block diagram.
Acceptable performance in normal operating conditions
may include distorted signals due to noise. On other hand
unwanted behaviours are caused by changes in the physical
structure of the system. Certain types of undesired behaviours
can be simple, changes in process parameters, as a result of
others; in this case it can be form of interference by actuators,
sensors, or internal structure of the system. Although it is
possible behavioural patterns resulting from large load change
in the power system network that lead to change in the system
parameters where are not explicitly identified or observed.
Another problem associated with the defective instruments
can be detected by comparing the output signal of the system
itself with the reference system. It is now possible to define
two types of unacceptable behaviours, dysfunction and erratic
behaviour. Dysfunction operation is caused by changes in
process parameters and corrected by the controller, which are
processed by change the control parameter in order to
overcome the fault. However, the erratic behaviour must be
diagnosed to find the defective part then corrected by the
This stage is based on manually synthesized FLC
architecture with two inputs as the error (Δω) and the change
in error (Δώ), while the output (Δu) is directed to the
excitation voltage loop driver then to the generator winding as
shown in Figure 3. Three membership functions were selected
for each of the input variables, while the output was selected
as a linear function since the inference engine used is a TSK
type [17].
Synchronous
Machine
Regulator
regul
Exciter
Synchronous
Machine
Power
System
Excitation
Excitation Control
Figure 3. : Block diagram of excitation control system.
A. Manual Tuning of the Scaling Factors
In the proposed fuzzy power system stabilizer, FPSS has
two inputs and single output, which means three scaling
factors are considered: the error (Ke), change of error (Kc)
and output (Ko). To improve the FLC response, the FLC
scaling factors were manually tuned. The scaling factors for
the first generator FLC1 are Ke1 for the error gain, Kc1 gain
for the change of error and Ko1 is the output gain. While the
second generator FLC2 are Ke2 is the error gain, Kc2 is the
gain for change of error and Ko2 is the output gain. The best
values established for the scaling factors are as shown in
Table I.
experiment trained FLCs based on the three-phase fault
condition was better than the single-phase fault condition.
TABLE I. MANUAL TUNING OF THE FLC SCALING FACTORS.
FLC1
FLC2
Ke1
Kc1
Ko1
Ke2
Kc2
Ko2
Manual
Tuning
2
3.75
2.25
5
3.75
10
Sample result of the manually tuned FLCs in comparison to
the M.B. stabilizer is shown in Figure 4 which illustrates the
grid power (Vm) in pu of the whole system after Static Var
Compensator (SVC) in normal operation and without PSS
stabilizer. The system response of FLCs controllers is slightly
better than the MB controller, whereas the system without
PSS controllers became unstable. The main reason for
simulating this stage is to observe the effect of the proposed
FLC controller on the system as a major step in the design
stages. But the main target is to focus on next stages which
are dealing with training of FLC using the Adaptive NeuroFuzzy Inference System (ANFIS), and auto tuning of the
scaling factors using SPSO for both controllers (FLC1 &
FLC2). In a later stage, a larger network contained four
generators and four controllers (FLC1, FLC2, FLC3 & FLC4)
working in the process at the same time.
1.05
without PSS
M.B. PSS
FLCMT PSS
1.04
1.03
Vm (pu)
1.02
1.01
1
0.99
0.98
0.97
0.96
0
1
2
3
4
5
6
Time (sec)
Figure 4. System’s response normal operation w/out PSS and with M.B and
FLC PSS.
B. FLC Training
In order to increase the controller response quality, the
FLC was trained using a learning signal form the MB
stabilizer using the ANFIS architecture [15]. The training is
performed in two steps, simulation with disruption in the grid
network by the occurrence of short circuit between one phase
and the ground, the next fault is a short circuit occurrence
between the three phases and the ground for a period of time
of 0.1 msec, as shown in Figure 5. Both trained controllers
were saved for each generator (FLC1 and FLC2). By the
Figure 5. FLC training in ANFIS editor.
Sample result of the manually tuned FLC in comparison
to the MB stabilizer is shown in Fig. 4 which illustrates the
grid power (Vm) in pu of the whole system after Static Var
Compensator (SVC) during fault in three-phase without PSS
stabilizer. The response of the manually tuned FLC is slightly
better than the MB controller, whereas the system without
PSS controllers became unstable. The main reason for
performing this stage is to observe the effect of the proposed
FLC controller on the system as a major step in the design
stages. But the main target is the focus on the next stages
which are dealing with training of FLC using the Adaptive
Neuro-Fuzzy Inference System (ANFIS), and auto tuning of
the scaling factors using PSO for both controllers (FLC1 &
FLC2) of the process at the same time.
C. Auto-Tuning FLC
In this stage the FLC scaling factors are selected using the
SPSO optimizer [18], which is aimed at improving the
response of both controllers keeping in mind the main
objectives. The objectives are dependent on the type and
requirements of the system to be controlled. For the power
system, the objective was to minimize the following four
variables on the final output (Vm) after the Static Var
Compensator (SVC) of whole system:



Minimizing the settling time.
Minimizing steady state error.
Minimizing the overshoot
Minimizing the first negative peak
IV. RESULTS ANALYSIS
A. Scaled-up system (four generators)
For the purpose to testing the efficiency of the new
designed system, it has been upgraded to a larger electrical
network grid, by using MATLAB Simulink Power System
Tools, which includes in this case four generators; two with a
B.
Three phase fault
As well as to test the supervisory control, how to
response for major interruption, such as three phase fault, two
fault breakers connected in the network. One of the breakers
closes after 4.8 sec of the start of the simulation for a
transition time period of 0.1 sec. Four neuro-fuzzy logic
controllers were used to stabilise each turbine. The FLCs
replace the conventional MB stabilizer to improve stability, as
shown in Figure 6. Same as before the controller was tuned
using SPSO to optimize the selection of the scaling factors.
The optimizer’s objective function is based on three
objectives: steady state error, settling time, overshoot, and the
negative peak. The first and second objectives have the
highest priority, while the third has medium priority, whereas
the fourth objective has the least priority.
M1
1000 MVA
13.8 KV/500 KV
D/Yg
M2
500 KV/13.8 KV
5000 MVA
Yg/D
Line 2
350 Km
Line 1
350 Km
The FLCs were trained on data generated using 3-phase
fault conditions which dubbed as 3-phase training. Four FLCs
controllers were trained for the generator. The scaling factors
were tuned automatically for each controller using SPSO
optimiser.
The system was tested and simulated for different fault
conditions, namely, single and multi-phase conditions.
Simulation results show that controllers have performed well
for single-phase fault, two and three-phase in comparison to
the performance when the system was driven without PSS.
Figures 7, 8 and 9 show the response of the system to the
faults when was driven by without PSS, MB stabilizer and
auto-tuned FLCs respectively.
Figure 7 shows the response of the FLCs controller with 3phase training and auto-tuning scaling factors to four
controllers. The first test was conducted for one phase fault,
and assumes that the settling time for the system at ±1.2 %.
1.1
w./out PSS
M.B.PSS
FLCAT PSS
1.05
Vm (pu)
capacity of 5000 MW, and the other two with a capacity of
1000 MW, all the generators are turbine driven. The four
generators mutually connected to the network via high
voltage the transformer, bas bar and transposed line with
length 1,500 km together with a load of 11,000 MW.
1
0.95
B1
B2
100 MW
B3
SVC
200 Mvar
5000 MW
0.9
400 MW
Line 1
100 Km
0
M3
1000 MVA
B1
500 KV/13.8 KV
Yg/D
Line 2
350 Km
Line 1
350 Km
100 MW
B2
M4
5000 MVA
B3
SVC
200 Mvar
5000 MW
Figure 6. Block diagram of the scaled-up power system.
The optimized values of the scaling factors for both the
auto-tuned fuzzy logic controller and the manually tuned are
listed in Table III.
TABLE III. FINAL AUTO-TUNING SCALING FACTOR VALUES.
FLC1
Ke1
FLC2
Kc1
2
4
6
8
10
12
Time (sec)
400 MW
13.8 KV/500 KV
D/Yg
Ko1
Ke2
Kc2
Ko2
Manual
Tuning
2
3.75
2.25
5
3.75
10
Auto.
Tuning
1.100
1.948
6.559
2.096
27.57
5.666
FLC3
FLC4
Manual
Tuning
2
3.75
2.25
5
3.75
10
Auto.
Tuning
1.139
0.941
9.215
1.556
6.006
5.543
Figure 7. System’s response to 1-phase fault with 3-phase training and Autotuning.
The fault start at 4.8 sec and end at 4.9 sec, the FLCs
controllers have reduced the stability time from 3.77 sec to
2.5 sec moreover with two peaks fluctuation compared to the
MB controllers with six peaks and larger overshoot such that
FLCs has 3.2% while MB has 3.7%. In Figure 8, the FLCs
have reduced the stability time from 4.03 sec to 2.43 sec, with
less fluctuation from six to two compared to MB controllers.
More importantly when a three-phase fault is simulation as
shown in Figure 9, the FLCs have reacted with high
efficiency compared to the MB and reduce the overshoot
from 4.9% to 4.2% and stability period from 3.41 sec to 2.5
sec with respect to the MB controller also in long period
started appearance of increasing on steady state error. The
numerical results are shown in Table IV.
C. Seven consecutive serious fault
In order to test the supervisory control system more
accurately, a scenario has been applied on the system process,
which include connect and lose a large load at seven deferent
place from the electrical network in sequence. The behaviours
of the all system are compared to its behaviours when was
governed by the other advanced conventional control systems.
Figure 8 show a comparison between the latest conventional
stabiliser controller and the supervisory power stability
control system for different faults. The seven different types
of faults during 30 second duration, while the first fault in the
scenario simulates a loss of 75 MW load at second 5 then
returned back to normal at second 8; at second 10, a cross
connection link was disconnected, and then went back to
normal at second 15; at the 20th second a high load loss
occurred with 400 MW capacity then returned to normal at
second 25. In the figure, it is shown that the supervisory
controller reaction is much smoother, stable, faster response
and less steady-state error compared to other competitive
controllers. Table IV shows the auto-tuned value of scale of
factors for all low level controllers, which are reset it by the
supervisory controller automatically concerning to the fault
type.
TABLE IV. SIMULATION RESULTS FOR MULTI FAULT WITH AUTO-TUNE AND
M-TRAINING.
Multi Band Controller (MB)
Settling
time (sec)
Overshoot
(%)
Fuzzy Logic Controller (FLCAT)
Settling
time (sec)
Fluctuation
Overshoot
(%)
Fluctuati
-on
1 phase
fault
3.77
3.7
6
2.5
3.2
2
2 phase
fault
4.03
3.7
6
2.43
4.4
2
3.41
4.9
6
1.15
FLC2
Ke1
Kc1
Ko1
Ke2
1.0823
2.6461
7.8087
1.9998
27.5140
5.1521
Load 1 with
1.1505
125 MW
1.9482
4.9597
2.0969
23.5773
5.6669
Load 2 with
1.7823
100 MW
1.9461
7.8087
1.8998
5.5140
6.5521
Load 5 with
1.0952
450 MW
1.9406
7.9116
2.0523
32.1371
5.6305
Network
separation
1.3467
2.2743
11.009
8.7224
5.8570
5.7979
Load 3 with
1.1393
125 MW
0.9416
9.2151
1.5564
2.0068
5.5439
Load 4 with
1.7805
100 MW
1.9948
7.8161
1.8986
5.5144
6.5832
Load 6 with
1.0906
450 MW
0.8479
8.5644
1.5054
11.9482
5.5334
Normal
Operation
2.5
4.2
Normal
Operation
2
w/ Generic Controller
w/ Advanced Power Stability System Controller
w/ Supervisory Power Stability System Controller
1.1
Vm (pu)
FLC1
FLC3
3 phase
fault
1.05
1
0.95
0.9
TABLE IV. FINAL AUTO-TUNING SCALING FACTOR VALUES FOR ALL
CONTROLLER IN DEFERENT SCENARIO.
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
Kc2
Ko2
FLC4
1.4805
1.9948
7.916
1.598
3.814
5.583
Load 1 with
1.1393
125 MW
0.9416
9.215
1.556
2.006
5.543
Load 2 with
1.7805
100 MW
1.9948
7.816
1.898
5.514
6.583
Load 5 with
1.0906
450 MW
0.8479
8.564
1.505
11.948
5.534
Network
separation
1.1041
1.3176
18.156
15.682
6.596
21.442
Load 3 with
1.1505
125 MW
1.9482
4.959
2.096
23.577
5.6669
Load 4 with
1.7823
100 MW
1.9461
7.808
1.899
5.5140
6.5521
Load 6 with
1.0952
450 MW
1.9406
7.911
2.052
32.137
5.6305
30
Time (sec)
Figure 8: Different fault types.
V. CONCLUSION
To conclude, it was shown that the supervisory power
stability control system (SPCS) has achieved better
performance compared to the conventional power stabiliser
controller, in particular when the SPCS was tuned with the
aid of SPSO to select the scaling factors for each controller
automatically. This type of control is far better than the
standard stabiliser during normal and fault conditions.
REFERENCES
[1] J. Rasmussen, "The role of hierarchical knowledge representation in
decision making and system management," IEEE Transactions on
Systems, Man and Cybernetics, vol. 15, pp. 234-243, 1985.
[2] C. W. De Silva and A. G. J. MacFarlane, "knowledge based structure for
robotics manipulators," 1988 Proc. IFAC. Workshop on Artificial
Intelligence in Real Time Control, pp. 143-148, 2007.
[3] L. A. Zadeh, "Fuzzy sets," Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems:
Selected Papers by Lotfi A.Zadeh, pp. 19-34, 1996.
[4] M. F. Abbod, D. G. von Keyserlingk, D. A. Linkens and M. Mahfouf,
"Survey of utilisation of fuzzy technology in medicine and
healthcare," Fuzzy Sets Syst., vol. 120, pp. 331-349, 2001.
[5] W. J. Kickert, "Towards an analysis of linguistic modelling," Fuzzy Sets
Syst., vol. 2, pp. 293-307, 1979.
[6] J. Kiszka, M. Kochanska and D. Sliwinska, "The influence of some
parameters on the accuracy of a fuzzy model," Industrial Applications
of Fuzzy Control, pp. 187-230, 1985.
[7] K. Matsushima and H. Sugiyama, "Human operator's fuzzy model in
man-machine system with a nonlinear controlled object," Digest
Industrial Applications in Fuzzy Control, pp. 175-185, 1985.
[8] R. M. Tong, "Synthesis of fuzzy models for industrial processes-some
recent results," International Journal of General System, vol. 4, pp.
143-162, 1978.
[9] R. Tong, "The evaluation of fuzzy models derived from experimental
data," Fuzzy Sets Syst., vol. 4, pp. 1-12, 1980.
[10] M. Sugeno and K. Tanaka, "Successive identification of a fuzzy model
and its applications to prediction of a complex system," Fuzzy Sets
Syst., vol. 42, pp. 315-334, 1991.
[11] C. Xu and Y. Lu, "Fuzzy model identification and self-learning for
dynamic systems," Systems, Man and Cybernetics, IEEE Transactions
On, vol. 17, pp. 683-689, 1987.
[12] F. Van Der Rhee, van Nauta Lemke, Hans R and J. G. Dijkman,
"Knowledge based fuzzy control of systems," Automatic Control,
IEEE Transactions On, vol. 35, pp. 148-155, 1990.
[13] C. G. Moore and C. Harris, "Indirect adaptive fuzzy control," Int J
Control, vol. 56, pp. 441-468, 1992.
[14] A. Sallama, M. Abbod and P. Turner, "Neuro-fuzzy system for power
generation quality improvements," in Universities Power Engineering
Conference (UPEC), 2012 47th International, 2012, pp. 1-6.
[15] A. Afzalian and D. Linkens, "Training of neurofuzzy power system
stabilisers using genetic algorithms," International Journal of
Electrical Power & Energy Systems, vol. 22, pp. 93-102, 2000.
[16] H. Bevrani and P. R. Daneshmand, "Fuzzy logic-based load-frequency
control concerning high penetration of wind turbines," Systems
Journal, IEEE, vol. 6, pp. 173-180, 2012.
[17] T. Hong and C. Lee, "Induction of fuzzy rules and membership
functions from training examples," Fuzzy Sets Syst., vol. 84, pp. 33-47,
1996.
[18] Z. Zhan, J. Zhang, Y. Li and H. Chung, "Adaptive particle swarm
optimization," Systems, Man, and Cybernetics, Part B: Cybernetics,
IEEE Transactions On, vol. 39, pp. 1362-1381, 2009.