Locating Sources of PQ disturbance using an Artificial Neural Network

1
Locating Sources of PQ disturbance using an
Artificial Neural Network

• Edward Bentley
• Director of Studies Ghanim Putrus
• Second supervisors
• Peter Minns
• Steve McDonald New and Renewable Energy Centre

2
Locating Sources of PQ disturbance using an
Artificial Neural Network
Presentation Outline

• Introduction
• Importance of Power Quality Monitoring
• PQ Events
• Existing approaches to location
• FFT Analysis
• Feature Vectors
• SOM
• Progress so far
• Conclusion

3
Locating Sources of PQ disturbance using an
Artificial Neural Network

• In modern power networks, the issue of electrical
  Power Quality (PQ) is becoming very important.
• This is due to-
• Continuous increase in using power electronic
  devices that draw current which is not
  sinusoidal creating a voltage distortion which
  affects all loads connected to the network.
• Increasing penetration of loads which are
  sensitive to such voltage disturbances, such as
  Personal Computers.
• As a result there is an increasing need for PQ
  to be monitored to establish the type, source and
  location of the disturbance, allowing remedial
  measures to be taken.

4
A mixture of power electronics and resistive
loads may be ok
5
Too many power electronic loads within a system
may interract causing malfunctioning
6
Current Harmonics for 130 VA ac/dc Switch-
Mode-Power-Supply (as found in PCs)
Hence, there is a need for PQ monitoring and
measurement of harmonics in order to ensure
proper functioning of equipment
7
Disturbances oscillatory transient
8
Disturbances impulsive transient

• x axis
  time(s) y axis Voltage (pu)

9
Disturbances sag
10
Disturbances swell
11
Disturbances DC offset
12
Disturbances Voltage Flicker

• 
  
  y axis V x axis time(s)

13
Disturbances Voltage Notching

• 
  x axis Voltage V, y axis
  time(s)

14
How do you locate a disturbance?
15
EXISTING TECHNIQUES

• In 2005, a technique was suggested that uses a
  combination of DWT, a supervised and an
  unsupervised Neural Network to successfully
  determine which of two network capacitors had
  been switched.
• In a power system, a bus is a heavy gauge
  conductor forming an electrical node. Only a very
  rudimentary system could be coped with,
  comprising 2 busses only.

16
EXISTING TECHNIQUES

• In 2007, another research achieved good accuracy
  (98) in locating capacitor switching transients
  using Wavelet Transform measurements and a hybrid
  Neural Network based on an 18 bus network, but a
  minimum of 4 sets of separate PQ monitors were
  required.
• Selection of the composition of the chosen
  feature vectors allowed accuracy to be achieved
  with a reasonable processing time

17
EXISTING TECHNIQUES

• In 2007, it proved possible, using voltage and
  current measurements, to establish whether
  capacitor switching was occurring upstream or
  downstream of the monitoring point using
  measurements made at a single location.
• Only one monitoring point was used, but no
  location at a single bus level was achieved.

18
Observations

• Actual components in a real power
• Real systems are not ideal, but
• Possess, inter alia,
• inductance and resistance

19
(No Transcript)
20
???

• Can one take advantage of the systems actual
  (non ideal) properties to locate the source of a
  PQ disturbance?

21
FFT

• The FFT analyzer is a batch processing device
• That is it samples the input signal for a
  specific time interval collecting the samples in
  a buffer,
• After which it performs the FFT calculation on
  that "batch" and
• Displays the resulting spectrum showing the
  magnitude, phase and frequency of the signal
  components.

22
FFT

• To find out whether a sampled signal contains a
  certain frequency
• Add up the consecutive samples multiplied by
  weights, positive when the weighting function is
  in the first half of its period and negative when
  its in the second half.
• In Fourier Analysis, the weighting function is a
  continuous sinewave.
• To test for a particular frequency, use the sine
  wave of that frequency. The accumulated sum will
  be close to zero if the signal does not contain a
  given frequency.

23
FFT

• Using the FFT technique, a base weighting
  function is applied to the signal under test,
  then frequency multiples (harmonics) of the
  weighting function 2x 3x 4x 5x ......etc.
• This procedure allows analysis of the signal
  under test to determine the levels of the various
  harmonics within it.

24
Observation

• If you monitor at one bus, a particular applied
  disturbance will have different measured levels
  of Fourier harmonic amplitude, depending upon
  where the given disturbance occurs within a
  system, owing to the presence of system reactances

25
Example

• In a simulation of the IEEE 14 bus system, for a
  given disturbance caused by switching a capacitor
  at bus 3, measured at bus 6 the following
  harmonic levels (v) were measured
• Second third fourth fifth sixth
  seventh
• 0.39 0.37 1.32 0.28 0.38
  0.08
• Switching at bus 4 again measured at bus 6 gives
  the following measurements-
• Second third fourth fifth sixth
  seventh
• 3.48 1.81 0.85 0.66 0.34
  0.69

The harmonic structure of a signal, monitored at
a given location, varies with its source
26
Observation

• If you monitor the magnitude of differing
  frequencies, for a given disturbance a feature
  vector can be obtained,
• The components of the vector varying depending
  upon the source of the disturbance in the system

27
Monitoring to create Feature Vectors
28
Monitoring to create Feature Vectors

• If monitoring at two locations is used, a
  combined feature vector may be obtained, giving
  greater power of identification
• For instance, monitoring a given disturbance
  (originating from bus 4), at bus 6 gave the
  following harmonics
• Second third fourth fifth
  sixth seventh

3.48 1.81 0.85 0.66
0.34 0.69
29
Monitoring to create Feature Vectors

• Monitoring the same given disturbance
  (originating from bus 4), at bus 8 gave the
  following harmonic measurements-
• Second third fourth fifth
  sixth seventh

2.47 0.72 2.20 0.80
0.88 0.20
30
Monitoring to create Feature Vectors

• Combined feature vector for disturbance
  originating at bus 4
• Measured at bus 6
• B6SEC B6THIRD B6FOUR B6FIV
  B6SIXTH B6SEV
• 3.48 1.81 0.85
  0.66 0.34 0.69
• Measured at bus 8
• B8SEC B8THIRD B8FOUR B8FIV B8SIX
  B8SEV
• 2.47 0.72 2.20
  0.80 0.88 0.20

31
Proposal

• Differing feature vectors should allow
  differentiation of source locations
  ...............
• HOW?

SELF ORGANISING MAP (SOM)
32
SELF ORGANISING MAP (SOM)

• SOM can organise incoming feature vectors so that
  input vectors which are topologically close to
  others in the input to the system appear so
  displayed in the output.
• The output forms a map of the feature vectors,
  often in 2 dimensions

33
SELF ORGANISING MAP (SOM)

• Big advantage
  
• Similar feature vectors are located adjacent to
  each other on the SOM.
• Feature vectors originating from adjacent
  locations in a power system will appear close to
  each other on the SOM

34
SELF ORGANISING MAP (SOM)

• SOM will locate feature vectors similar to those
  it is trained with, and locate them in an
  appropriate location.
• You can train the system using signals from
  defined busses, and the system can interpolate
  the location of signals originating between
  busses
• A SOM normally comprises a 2-dimensional grid of
  processing elements known as nodes, operated in
  computer software.
• A model of the data representing a measurement is
  associated with each node

35
THE PERCEPTRON
36
ARCHITECTURE OF SOM
37
UPDATING WINNER NEURON AND ITS NEIGHBOURS
38
Phonemes Represented in an SOM

• Each node in this grid holds a model of a
  short-time spectrum derived from natural speech.
• Neighbouring models are mutually similar
• The SOM algorithm deals with the models in such a
  way that they recreate the topology of the
  observations

39
Application

• A different map will be required for each class
  of disturbance to be located
• The final system will identify a PQ disturbance
  using existing technique to enable the correct
  map to be used

40
(No Transcript)
41
PRACTICAL WORK

• IEEE 14 bus system modelled in PSCAD software
• 10,000 uF capacitor switched at each bus in turn
• Harmonic components 2nd to 7th recorded at buses
  6 and 8 using FFT
• Combined feature vectors obtained

42
Feature Vector produced from disturbance at Bus 8
43
Full Array of Feature Vectors made from
disturbances at all Busses
44
COMPLETED SOM SHOWING BUS LOCATIONS
45
FEATURE VECTOR FROM BUS 12 APPLIED TO SOM WITH
ERRONEOUS LOCATION
46
U-MATRIX DIAGRAM
47
NORMALISED RESULTS

• 12
• n B6SEC B6THIR B6FOUR B6FIV B6SIX B6SEV B8SEC
  B8THIR B8FOUR B8FIV B8SIX B8SEV
• 17.1 4.85 10.52 5.12 4.85 7.58 12.45 10.38 15.3
  6.04 2.83 3.02 Bus8
• 14.4 20.0 6.2 0.2 5.6 3.6 18.5 7.13 8.6 3.6 5.47
  6.67 Bus13
• 13.04 4.19 18.42 3.71 5.14 5.50 17.2 9.04 13.77
  3.53 6.02 0.43 Bus12
• 15.2 6.47 12.18 4.7 5.96 5.46 18.95 7.33 8.48
  3.62 5.44 6.18 Bus14
• 15.44 2.94 19.85 4.99 5.15 1.76 22.57 9.1 6.46
  5.26 4.56 2.04 Bus9
• 19.87 3.77 14.89 6.03 0.62 4.82 20.05 7.1 9.79
  5.49 4.3 3.28 Bus10
• 12.82 10.75 14.09 5.98 1.93 4.44 23.65 4.8 10.68
  5.68 2.5 2.5 Bus11
• 10.92 5.29 15.02 6.31 4.44 8.02 17.19 6.36 12.23
  4.10 4.67 5.45 Bus6
• 15.91 7.19 9.71 5.32 4.27 7.60 10.69 12.76 14.14
  6.21 2.41 3.79 Bus5
• 7.64 10.19 14.61 7.10 3.35 7.10 16.80 8.92 14.25
  3.18 6.21 0.637 Bus1
• 10.10 10.0 11.53 3.78 6.33 8.27 14.83 4.03 13.75
  7.37 2.95 7.07 Bus2
• 6.99 6.58 23.43 4.93 6.71 1.37 24.09 1.11 11.93
  6.56 1.74 4.58 Bus3
• 22.16 11.56 5.40 4.24 2.21 4.43 17.01 4.98 15.14
  5.5 6.02 1.35 Bus4
• 12.5 5.0 12.5 5.0 6.25 8.75 26.46 10.83 7.13 3.65
  0.625 1.30 Bus7

48
NORMALISED U-MATRI X
49
CORRECT IDENTIFICATION OF SIGNAL FROM BUS 7
50
PRELIMINARY RESULTS

• Som very sensitive to normalisation of signal
  amplitudes
• After due attention to this point 13/14 busses so
  far correctly identified using 2 monitoring
  points
• 8/14 busses correctly identified
• using 1 monitoring point only
• Small changes made to a feature vector give
  small change in location as expected.

51
Further Improvements

• 100 location accuracy using 2 monitoring points
  achieved when already normalised feature vectors
  have variance normalised over the full data set
  used to train the SOM-
• Once normalised feature vector -
• 10.92 5.29 15.02 6.31 4.44 8.02
  17.19 6.36 12.23 4.10 4.67 5.45
  0 0
• After second normalisation-
• -0.69 -0.553 0.317 0.916 -0.025 1.012
  -0.316 -0.346 0.226 -0.657 0.384 0.89 0
  0

52
Monitor Current in 2 Cables to Bus 10
53
1 Bus Measurement

• 100 location accuracy achieved with monitoring
  at 1 bus only measuring 2 separate cable currents

54
General Application

• The system has been tested and found satisfactory
  with a number of fault conditions including
• Oscillatory transient
• Sag
• Swell
• Harmonic distortion

55
PROBLEMS
56
Location Error for Bus 11 With 25 Disturbance
Power
57
Attempted Location of Sag at Bus 4 With Harmonic
Interference
58
Future Work

• Develop software to transfer data from PSCAD
  environment to MATLAB for analysis
• Improve robustness of technique
• Develop SOM for each type of disturbance
• Develop location system for disturbances not on
  system busses
• Combine the new system with an existing
  technique, based upon DWT measurements of a
  signal, thus allowing the creation and invocation
  of suitable initial conditions and the correct
  SOM mapping for the PQ event concerned.
• Build real system and test using DSP sampling
  techniques

59
Acknowledgements-PQ Event Diagrams taken from-