Transcript No Slide Title

Neural-Fuzzy Pattern Recognition
Algorithm for Classifying the Events
in Power System Networks
Slavko Vasilic
Department of Electrical Engineering
Texas A&M University
Outline
•
•
•
•
•
•
•
•
Problem, Goal, Objectives
Protective relaying
Neural network (NN) algorithm
Process modeling and simulation
Algorithm implementation
Fuzzyfication of NN outputs
Algorithm Testing
Conclusion, Future Work
Problem
•
Traditional relay settings are computed
ahead of time based on worst case fault
conditions and related phasors
•
The settings may be incorrect for the
unfolding events
•
The actual transients may cause a
measurement error that can cause a
significant impact on the phasor estimates
Goal
•
Design a new relaying strategy that does not
have traditional relay setting
•
Optimize the algorithm performance in each
prevailing network conditions
•
Improve simultaneously both, dependability
and security of the relay operation
•
Demonstrate the benefits using realistic
network and fault events
Objectives
•
Implement a new pattern recognition based
protection algorithm
•
Use a neural network and apply it directly to
the samples of voltage and current signals
•
Produce the fault type and zone
classification in real time
•
Study various approaches for preprocessing
NN inputs and fuzzyfication of NN outputs
Protective Relaying
The different parts of the fault clearance chain
bus
Bre ak e r
e le m e nt
Bre ak e r
m e chanis m
Trip coil
line
CT
Dis tance
Re lay
VT
Auxilary
Re lay
Protective Relaying
The principle of distance protection relays
t
t3
t2
t1
C
A
ZA
F1
F2
ZB
ZC
l
F3
ZD
t
t3
t2
t1
l
B
D
Protective Relaying
Mho fault characteristic of distance relay
X
line impedance
locus
zone 3
impedance trajectory
during fault
zone 2
zone 1
LOAD
AREA
R
Neural Network Algorithm
The principle of multilayer neural networks
NEURAL NETWORK
x0
HIDDEN
LAYER OF
NEURONS
NEURONS

Target
OUTPUT
LAYER OF
1
x1
z0
z1

1
+
y1
Input
Output
xi
wji
j

zj
wkj

-
k

yk
Error
xM
N

zN

L
Adjust weights and
numb er of hidden neurons
yL
Neural Network Algorithm
Pattern classification of faulted events
Patterns
Class
decision
boundaries
Neural Network Algorithm
Characteristic of the used neural network
•
•
•
•
•
•
Direct use of samples (no feature extraction)
Hidden layer of competitive neurons
Self-organizing
Unsupervised and supervised learning
Outputs are prototypes of typical patterns
Adaptability for non-stationary inputs
Neural Network Algorithm
Training steps
Start
Patterns Normalization
UNSUPERVISED
LEARNING
New Iteration
Initialization Phase
Stabilization Phase
SUPERVISED
LEARNING
Clean and Mixed Clusters
Identification
Clean Clusters Class
Membership Assignement
Clean Clusters Extraction
Input Set Reduction
Decreasing Vigilance
Parameter 
Number of Mixed
Clusters is Zero
or  <  ?
Yes
End
No
Process Modeling and Simulation
RE HL&P Stp-Sky power network model
Process Modeling and Simulation
Scenario cases: general fault events
•
•
All types of fault (11 types)
•
•
Fault impedance variation (0-100 Ohms)
Fault location variation (0-100% of the line
length)
Fault inception angle variation (0-360 deg)
Process Modeling and Simulation
Example of patterns for various fault parameters
Algorithm Implementation
Training and testing
•
Power network model is used to simulate
various fault events
•
Fault events are determined with varying
fault parameters: type, location, impedance
and inception time
•
The simulation results are used for forming
the inputs for algorithm training and
evaluation
Algorithm Implementation
Training and testing (cont’d)
•
•
•
Training tasks are aimed at recognizing fault
type and location
Test patterns correspond to a new set of
previously unseen scenarios
Test patterns are classified according to their
similarity to the prototypes by applying
K-nearest neighbor classifier (decision rule)
Algorithm Implementation
Properties of input signal processing
•
Data selected for training: currents, voltages
or both
•
•
•
•
Sampling frequency
Moving data window length
Analog filter characteristics
Scaling ratio between voltage and current
samples
Algorithm Implementation
Moving data window for taking the samples
Algorithm Implementation
Example of the patterns for various scaling ratios
Algorithm Implementation
Prototype
Training
patterns
Algorithm Implementation
The outcome of training are pattern prototypes
Fuzzyfication of NN Outputs
Fuzzyfied classification of a test pattern
Fuzzyfication of NN Outputs
Fuzzyfied classification of a test pattern
•
Determine appropriate number of nearest
prototypes to be taken into account
•
Include the weighted distances between a
pattern and selected prototypes
•
Include the size of selected prototypes
Fuzzyfiaction of NN Outputs
Fuzzyfication of NN Outputs
Fuzzy K-NN parameter optimization
Algorithm Testing
Test
pattern
Nearest
prototypes
Algorithm Testing
Propagation of classif. error during testing
Algorithm Testing
Algorithm sensitivity versus data used for training
Fault type and zone classification
Classif. error [%]
10
8
9
Basic K-NN
Fuzzy K-NN
7.6
6
3.8
4
2.7
2
2
1.4
0
Voltages
Currents
Currents and
Voltages
Conclusion
•
Protection algorithm is based on unique selforganized neural network and uses voltages
and currents as inputs
•
Tuning of input signal preprocessing steps
significantly affects algorithm behavior
during training and testing
•
Fuzzyfication of NN outputs improves
algorithm selectivity for previously unseen
events
Conclusion
•
The algorithm establishes prototypes of
typical patterns (events)
•
Proposed approach enables accurate fault
type and fault location classification
•
The power network model is used to simulate
a variety of fault and normal events
Future Work
•
Perform comprehensive algorithm training for
extended set of training patterns
•
Perform extensive algorithm testing and
performance optimization
•
Study algorithm sensitivity versus various
input signal preprocessing steps
•
Implement algorithm on-line learning