A FUZZY INFERENCE SYSTEM TO VOLTAGE/VAR CONTROL IN DMS -

DISTRIBUTION MANAGEMENT SYSTEM

Vladimiro Miranda
[email protected]
Patricia Calisto
[email protected]
INESC - Instituto de Engenharia de Sistemas e Computadores do Porto
P. Repblica 93 - 4050 Porto - PORTUGAL - Fax: 351.2.205 55 00
FEUP Faculdade de Engenharia, Universidade do Porto

Abstract This paper demonstrates the feasibility of the
concept of fuzzy controller for voltage control or loss
minimization. Cascading Mamdani controllers are used
together with sensitivities derived after a fuzzy clustering
process over voltage profiles. Three to four iteration steps with
a Newton-Raphson power flow routine are enough to recover
voltages into admissible bands, demonstrating the efficiency of
the approach.
Keywords: Evolutionary Algorithms, Particle Swarm
Optimization, Voltage/Var control
1. INTRODUCTION
One of the modules needed in DMS (Distribution
Management Systems) or EMS (Energy Management
Systems) is one that suggests a number of control actions to
keep node voltages inside limits and move the operation
point so that losses are minimized. The control actions may
be exerted on transformer taps, capacitor banks, generator
excitation or FACTS devices. Distribution networks are at
present becoming very complex, namely with the addition
in large scale of distributed generation, and in some cases
this generation also allows voltage control.
In a DMS environment, data are gathered by a SCADA
and control actions exerted through it. These actions may
result automatically from software or may be issued by an
operator, after receiving advice from a decision aid tool.
In this paper, we will describe a new architecture for a
DMS module aiming at Voltage/Var control and loss
minimization. This architecture is based on the articulation
of cascading fuzzy controllers of the Mamdani type,
constituting a fuzzy inference system, with a power flow
routine that periodically evaluates the effects of control
actions. The fuzzy rules are related to the efficiency of a
control action and the set point of each control.
Besides aiming at keeping all voltages inside the allowed
band or reducing system losses, the fuzzy inference
controller also aims at retaining controllability of the
system, i.e., it suggests a control vector of actions that will,
as much as possible, keep the control set points well inside
their possible range.
Before a final control decision is reached, the controller
simulates a series of control steps and interacts with a
power flow routine to evaluate the consequences of each
move and successively correct and adapt the sensitivities
for control efficiency and control margin.
One will show that in a realistic problem with several
possible control actions (generator excitation and
transformer taps) the controller reaches a decision solution
in a short number of steps (demanding only 3 to 4 power
flow calculations) or iterations and therefore the technique
is suitable for real-time application in a DMS environment.
A previous and very interesting and successful work
must be referenced, with application in practice, following
the idea of building a fuzzy reasoning platform to organize
a voltage and reactive power controller [1], and in general
of introducing fuzzy set concepts in the control of power
systems and subsystems [2].
Most efforts in the past in the application of fuzzy
control techniques to power systems had been applied to
areas related with stability. Interesting and didactic
examples of these attempts are described in [3]
The approach of using fuzzy logic concepts in power
system control challenges the use of heavy alternatives,
such as using simulated annealing [4] or genetic algorithms
for the same purpose of deriving control actions to control
voltage levels or reduce losses, because it seems to be some
orders of magnitude more efficient in terms of computing
effort. The robustness of the fuzzy logic approach in
extremely difficult convergence conditions remains to be
researched however, this is also a challenge to other
approaches.
2. BASIC DATA AND VARIABLES
For simplicity, in this chapter we will describe the basic
data and variables for a Voltage controller, because a
controller aiming at reducing losses has the same structure.
In a DMS/EMS database, we will have available data
about lines and network topology, transformers (namely
tap-changing), generators or capacitor banks. The database
also stores data specified by the operator, such as limits for
over- and under-voltage at the nodes of the grid.
Furthermore, from the SCADA or from the State
Estimation module of the DMS/EMS one receives timely
information on loads and voltages, as well as the position of
controllers in the system.
14th PSCC, Sevilla, 24-28 June 2002 Session 03, Paper 1, Page 1

As control devices, we have assumed that one could act
on transformer taps, switching capacitor banks or generator
excitation. However, the model does not exclude other
forms of control.
The purpose was to design a controller that, in a
DMS/EMS environment, would react when sensing that
voltage thresholds were being violated. This reaction would
mean defining new set points for generator voltage or for
transformer (or capacitor) taps.
The factors to take in account are:
The evaluation of severity of the most violated
constraint the controller will act to eliminate the
most severe constraint violation at any time. For
voltages, this means that the corresponding node
must be identified
The efficiency of a control action to reduce a
constraint violation, one has available a set of
control actions (one can select any one out of a
number of tap-changing transformers or generators).
However, depending on their nature and the location
of the respective devices in the network, the impact
of each action will be different.
The availability for control of the control action
using a certain transformer tap to correct some low
voltage could be very efficient if only the
position of the controlling device were not already
at its limit.
These factors may all receive a linguistic description,
once one defines their universes of discourse.
2.1. Efficiency
Each change in a control position
j
(.) at a node j will
have impact on the voltage value at another node k. A
classical way to evaluate such impact is by means of
sensitivity coefficients S
jk
obtained from partial derivative
concepts
j
k
kj
V
S

=
However, partial derivatives are local concepts and do
not give a correct picture of the shape of the function and of
the real impact a in controlled value for larger than
infinitesimal changes. This has been well observed in [1],
where a scheme based on Experiment Planning has been
suggested to derive more adequate sensitivities.
In any case, it has been shown that sensitivities evaluated
from imposing variations in controls and checking the
impact on the nodal voltages work better. In our work we
did not follow the suggestion of [1] but we actually built
sensitivity matrices by organizing a series of experiments with
discrete step changes in each possible controlling variable.
The sensitivity values were scaled into an interval [-1,1]
generating therefore a signal called efficiency.
2.2. Fuzzy clustering of load profiles
A set of matrices may be organized and stored as data, to
be selected for use depending on the loading condition of
the power system. In [1] the method followed has been to
define three Load Conditions (Low, Medium and High
Load), to calculate the Sensitivities for each Load condition
and then, when facing an actual vector of nodal loads, to
calculate an interpolated value for the sensitivity factors.
Not far from such idea, the model in this paper suggests
the adoption of concepts from fuzzy clustering.
A Power Profile is a vector L(t) of nodal generations and
loads at a certain time t. Because different types of
consumers will be supplied at distinct locations, the load
will not rise or descend simultaneously at all nodes.
It is therefore conceivable that one may build a set of
power profiles, from several days at distinct hours, and then
apply a clustering technique in order to find some meta-
structure in such set.
We have selected the Bezdeks Fuzzy c-Means [5] fuzzy
clustering algorithm for this purpose. The main reasons are:
the method allows us to identify the centroids or
prototypes of the clusters defined, and
the method associates to each Power Profile a
membership value to each cluster, represented by its
centroid.
Then, given a Power Profile, it is possible to calculate its
membership to each cluster. Define
V
i
: vector denoting the prototype of cluster i
L(t): vector of the power profile at time t
d
i
: distance between L(t) and the prototype of cluster i,
with d
i
= ||L(t) V
i
|| (usually calculated with an
Euclidean norm)
c : number of clusters

i
: membership of profile L(t) to cluster i
[ [ + , 1 m : weighting exponent, a parameter defining
the more fuzzy or hard character of the
clustering; adopted values for m are close to 2.
We have then
1
c
1 k
) 1 m (
2
k
i
i
d
d

=

1
1
1
]
1

'
|
=

As we are dealing with a fuzzy partition, we know that

=
=
c
1 k
k
1
A sensitivity coefficient S(t) associated with a given
power profile at time t may then be calculated from the
corresponding sensitivity coefficients S
k
for the prototypes:

=
=
c
1 k
k k
S ) t ( S
14th PSCC, Sevilla, 24-28 June 2002 Session 03, Paper 1, Page 2

In practice, instead of performing a fuzzy clustering of
power profiles, one may directly define some anchor
profiles V
k
that will act as prototypes. For instance, one
may select profiles for Low, Medium and High Load,
calculate the sensitivities associated to them and store them
in the DMS/EMS data base to be used in calculations at any
time t for profile L(t).
2.3. Evaluation of severity
For Voltage control, the condition to be met is
Max
k
k
min
k
V V V for all k nodes.
This creates a dead band where no control action is
required. Outside this band, the severity of the violation is
proportional to its value, either
min
k
k
V V or
Max
k
k
V V .
A plausible interval for violations is defined, say [a,b]
and a violation in this interval is made to correspond to a
value in an interval [-1, 1]. We have, therefore, created a
signal called severity.
2.4. Position
A controlling variable
j
is associated with a control
range [ ]
Max
j
min
j
, and it may be changed within such
interval, which includes in a more or less central location
the nominal value
nom
j
.
We have mapped the interval [ ]
nom
j
min
j
, into
[ ] 1 , 85 . 0 and [ ]
Max
j
nom
j
, into [ ] 15 . 1 , 1 and therefore
built a signal called position of the control.
3. CONTROLLER DESIGN
3.1. Basic architecture
The Voltage controller must produce a control command
that is a vector of control signals, one for each controlling
device.
In our model, each of the controlling device signals is
controlled by a controller with the architecture displayed in
Figure 1.
The control system relies on two Mamdani fuzzy
controllers, C1 and C2. Together, they accumulate a rule
base; the activation of these rules generates the control
signal required for a control device (e.g., a generator
excitation or a tap changing transformer).
In both controllers we have used, for simplicity and with
good results, triangular fuzzy sets.
The universe of discourse of all variables has been
partitioned in 5 linguistic values NB (negative big); NS
(negative small); ZE (zero); PS (positive small); PB
(positive big).
position
efficiency
violation
Control signal
Control power
C1 - Controller 1
C2 - Controller 2

Figure 1 Cascading fuzzy controllers forming a control
block
These variables have been evenly distributed within the
interval defining the universe of discourse [-1, 1] in all
cases except for position.
In the fuzzy set operations performed, we have used the
min operator as the T-norm (for intersections and
implications) and the Max operator as the S-norm (for
union).
The defuzzification method adopted has been the Center
of Mass, in both fuzzy controllers. This means that the
output of C1 is defuzzified before inputting in C2.
The defuzzified control signal output, in the range [-1,
1], is then mapped into the admissible range for the
controlling device. If this range is composed of discrete
values (such as transformer taps), it is rounded to the most
convenient nearest value.
The fuzzy associative memory or rule map for C1 and
C2 has the form of Table 1.
Table 1 Rule map for the Fuzzy Controllers (meaning of
linguistic labels: in the text)
NB NS ZE PS PB
NB PB PB PB PS ZE
NS PB PB PS ZE NS
ZE PB PS ZE PS NB.
PS PS ZE NS NB NB
PB ZE NS NB NB NB

The controller C1 is designed as a proportional
controller and controller C2 is designed as derivative, i.e.,
its control signal represents a deviation to the previous
controlled variable value.
This design for C1 guarantees that a controlling variable
near its limits will only be used if this represents moving
the control variable to inside the interval and that there
14th PSCC, Sevilla, 24-28 June 2002 Session 03, Paper 1, Page 3

will be an attempt to move it inside whenever it is
beneficial.
Therefore, this design favors the maintenance of control
variables inside their ranges this means that the controller
system tends to keep a control margin available in every
variable, which is very important and useful from an
operation point of view.
3.2. Controller algorithm
The controller algorithm runs in a series of iterative
steps:
1. Evaluate the largest voltage violation
2. Evaluate the efficiency and position of available
control devices
3. Generate an incremental control signal that changes
control states (taps or generator excitation)
4. Limit changes by applying an iteration step size
5. Evaluate system voltages for the corrected profile
after applying the control signals. If violations still exist
go back to point 1, else stop.
The evaluation of system voltages is done using a power
flow routine (in our case, with a Newton-Raphson
algorithm).


Network
Network model
(power flow)
Fuzzy
Controller
Voltages Changes in control
variables

Figure 2 General scheme of the control loop

4. APPLICATION EXAMPLE
Following the principles described in the previous
sections, we have tested the concept over a system based on
a real network from Brazil, whose data are described in
reference [1].
The system is represented by a network with 26 buses,
including 5 power injection buses and 21 load buses see
Figure 3. It has 48 branches and 12 transformers, and the
total load is in the range of 800 MW and 350 MVAr for a
peak load profile.
There are two main voltage levels in the system, and the
main power injection comes from an interface with an
interconnected transmission system (bus 7). The other
power injection sources are generators.
The operator has the possibility to control voltage levels
by acting at 6 places: three generators and three transformer
groups (transformers are in parallel and, of course, their
taps are changed simultaneously).
Running a power flow routine shows that, at peak hour,
many node voltages are well below the limit of 0.9 p.u.
determined as the admissible threshold. In fact, 50% of the
buses had a voltage below that limit see Figure 4.
We have calculated the sensitivity coefficients for node
voltages relative to the 6 controlling variables, and built the
required matrices.



Figure 3 Single line diagram of the test system.
14th PSCC, Sevilla, 24-28 June 2002 Session 03, Paper 1, Page 4


0,80
0,90
1,00
1,10
13579
1
1
1
3
1
5
1
7
1
9
2
1
2
3
2
5
Tenses nos barramentos (p.u.)

Figure 4 Voltage profile at peak hour with controlling
variables at nominal value (no control action)


Figure 5 Voltage profile at peak hour after the control
action proposed by the fuzzy controller system

In order to build the voltage controller, we have
prepared 6 control blocks similar to the one in Figure 1 a
block to each control variable (3 for generator voltages and
3 for transformer taps).
All simulations were performed in MATLAB
environment and with a Newton-Raphson power flow
routine.
Applying the control algorithm, in only three iterations
we reached a fully feasible network condition, with all
voltages inside the admissible band see Figure 5. Also, no
voltage appeared above 1.15 p.u., which was set as the
upper limit.
This speed of convergence has been recognized in a
number of runs and experiments made. Typically, only
three to four iterations are required until a feasible voltage
profile is reached.
5. EXPERIMENT WITH AN ANN
We have conducted a further experiment, replacing the
fuzzy controller C2 by an artificial neural network (ANN)
representing controller C3. We have thus generated 1053
power profiles and generated a data set by applying the
fuzzy controller and recording the adequate control signal
to restore voltages to their admissible band.
This set has been divided into a training and a test set,
and a neural network has been trained. The architecture of a
successful ANN was 2-6-2-5-2-1, with linear activation
functions for the input and output layers and hyperbolic
tangent activation function for the neurons in the other
layers.
The structure of a control block is now the one
represented in Figure 6.
position
efficiency
violation
Control signal
Control power
C1 - Controller 1
C2 - Controller 3

Figure 6 Control block with an ANN replacing one of the
fuzzy controllers

As expected, when tested this scheme offered
comparable results of the same quality as the one using only
Mamdani fuzzy controllers.
In terms of practical implementation in a DMS/EMS
environment, the scheme using the ANN may offer some
slight advantages in terms of the generation and
organization of the software modules.
6. AVOIDING CYCLING
The model has proved well in recovering feasibility for
systems with violation of constraints. However, if the
algorithm cannot find a feasible solution, it may enter in a
cycle, alternatively correcting voltage at one bus, which
originates a violation at another bus whose correction, in
turn, will generate again a violation at the previous bus and
so forth.
To prevent this to happen and guarantee a solution, there
are two mechanisms available:
a) upon detection of cycling or after a number of
iterations without reaching a stable solution, the
14th PSCC, Sevilla, 24-28 June 2002 Session 03, Paper 1, Page 5

iteration step is reduced; under a given threshold
of step size, the algorithm is stopped
b) the best-so-far solution is memorized, to be
available when the process stops.
The best-so-far solution is measured in terms of the
nature of the constraint violations detected. For voltage
control, two ways of selection have been included in the
model: one that evaluates continuously the average value of
violations and another that monitors and implements a min-
max strategy, i.e., gives preference to minimizing the
maximum violation detected.
7. MINIMIZING LOSSES
The basic scheme to achieve loss minimization is similar
to the one presented for Voltage control. The following
main adaptations are necessary:
a) calculation of sensitivity matrices relating variation
in losses with control actions again, instead of
looking for partial derivatives, one looks for linear
relations between changes (by step increments) in
control variables and changes in system losses
b) redefinition of the variable violation for power
losses, a violation will be seen as the difference
between actual losses and a pre-determined target of
inferior value (could be 0, in an extreme case).
It goes without saying that the action of the voltage
controller, by itself, already contributes significantly to
reduce losses.
8. CONCLUSIONS
This paper demonstrates the feasibility of using a
cascading set of fuzzy controllers of the Mamdani type to
build a voltage control system to be integrated in a
DMS/EMS environment.
The model developed gives place to an extremely
efficient procedure that only requires a very small number
of power flow runs to reach a good solution.
As a first order approach, it demonstrates great
superiority over methods that search for optimal controls by
using heavy algorithms such as simulated annealing or
evolutionary computing. These ones will require hundreds
of evaluations using a load flow routine, while the Fuzzy
Control system only requires typically a number below 6.






However, because it is a method relying on sensitivities,
one cannot hide that there is the theoretical possibility it
may get trapped in local optima.
Of course, if one has available an alternative robust
optimization method, one may always be able to generate
off-line a training set of higher quality and try to improve
the training of Artificial Neural Networks to be used as
exemplified in Section 5 of this paper.
The report presented in this paper confirms therefore the
results obtained by other researchers that the application of
concepts of fuzzy inference to the Voltage/Var control and
loss minimization problems is an excellent approach.
Furthermore, one demonstrates that there is no need for a
complex model building a fuzzy rule base and an inference
engine: simply cascading Mamdani controllers achieves a
very good result.

REFERENCES

[1] Marcio Junges, Fuzzy Logic in voltage and reactive
power control in power systems (in Portuguese,
Lgica fuzzy em controle de tenso e potncia reativa
em sistemas de potncia ), M.Sc. Thesis, Pontifcia
Universidade Catlica de Minas Gerais, no.
621.31J95|T - PucMinas Library, Belo Horizonte,
Brazil, 2000.
[2] Petr Ekel, Luiz Terra, Marcio Junges, Vladimir Popov
Fuzzy Technology in the Design, Planning and
Control of Power Systems and Subsystems,
Procedings of EUFIT98 - European Congress on
Intelligent Techniques and Soft Computing, vol.2., pp.
1126-1130, Aachen, Germany, 1998
[3] Electric Power Applications of Fuzzy Systems
(book), ed. Mohamed El-Hawary, IEEE Press Series
on Power Engineering, 1998
[4] Jorge Pereira, J. Tom Saraiva, Maria Teresa Ponce de
Leo, "Identification of Operation Strategies of
Distribution Networks Using a Simulated Annealing
Approach", Proceedings of IEEE Budapest Power
Tech'99, paper BPT99-357-17, August 1999
[5] James C. Bezdek, Pattern Recognition with Fuzzy
Objective Function Algorithms (book), Plenum Press,
New York, 1981
14th PSCC, Sevilla, 24-28 June 2002 Session 03, Paper 1, Page 6