This document discusses methods for detecting and classifying disturbances in a hybrid distributed power system using wavelet transform and artificial neural networks. It begins by introducing the motivation for distributed generation and some challenges like unintentional islanding. It then describes using the wavelet transform on voltage signals to detect islanding and compares this to using total harmonic distortion. Statistical indices from the wavelet transform are used as inputs to an artificial neural network classifier to identify different disturbances like islanding, faults, and load changes with over 95% accuracy. The study concludes the wavelet transform approach provides better detection and classification than conventional techniques. Future work could explore improving performance under noisy conditions and different feature selection.

1. Detection and Classification
of Disturbances in a Hybrid
Distributed System Using
Wavelet Transform and ANN
GUIDE : Prof. P. R. Subadhra
By,
Sleeba Paul Puthenpurakel

2. Introduction
Rapid Increment in Energy Demand
India : By 2030, Demand = 6 x Current Demand
Feasibility of Renewable Energy Sources
Minimizes the environmental pollution
Advancement in Power Electronics, Internet of Things and Automation

3. Distributed Generation
Power Generation is no more oligopolistic
Advanced Reliability of Power
Types
Stand Alone
Grid Connected
DG Penetration Increment

4. Islanding

5. Intentional Islanding
Informed at Grid and DG side
Reduce congestion in network
Improves voltage profile
Separate the faulty part from network

6. Unintentional Islanding
Safety
Change in Impedance level
Protection Scheme
Ferroresonance
Over Loading of DG

7. Islanding Detection Methods
Communication Based
Communication between DG and Utility
Heavily relies on communication system
Reliable but complex and costly
Active
Perturb System Variables ( Voltage and Frequency)
Small Non - Detection Zone

8. Islanding Detection Methods
Passive
Measures system variables ( Voltage and Frequency )
No Power Quality Issues
Large non - detection zone
Selecting right variables is crucial
Threshold fixing is difficult

9. Computational Intelligence based methods
Mimics human intelligence
Solves non - linear multi objective problems
High speed and accuracy
Learn from examples
Training of algorithm is a one time process
Can detect and classify disturbances

10. Motivation of Thesis
Importance of Distributed Generation
Unintentional Islanding hazards
Inabilities of conventional methods
Power Quality issues
High Non – Detection zone
Cost of implementation
Popularity and Robustness of Computational Intelligence and Machine
learning

11. Objectives
Disturbance Detection using a conventional method in Hybrid System
Disturbance Detection using Wavelet Transform in Hybrid System
Disturbance Classification using Computational Intelligence Method
Comparative study

12. Hybrid System

13. Hybrid System Specifications
PV System
Rated Power - 250 kW
Irradiance - 1000 W/m2
Wind Power Plant
Rated Power - 1.5 MW
Wind Speed - 12 m/s
Grid

14. Conventional Islanding Detection Method
Detection of Harmonics Method
THD
Measure of Harmonic Content in the signal
Based on Fourier Transform
Passive Method
THD of Voltage Signal at PCC is calculated
Normal value of THD at grid connected mode < 5% ( IEEE Standard)

15. Transform of a Signal
Real world signals are time domain ( Time V/s Amplitude )
Transform gives additional information
Frequency domain
Stationary signal
Frequency content dont change with time
All frequencies are present in all time !!

16. Fourier Transform - Stationary Signal

17. Fourier Transform - Non Stationary Signal

18. Comparison of Frequency Responses
FT - STATIONARY SIGNAL FT - NON-STATIONARY SIGNAL

19. Fourier Transform - Anxieties
Gives the overall frequency content information
Frequency and Amplitude
Misses the time domain information
Can’t extract complete information from Non - Stationary signals
Frequency at a particular instant of time can’t be calculated
Can't differentiate events using the variation in frequency content of

20. THD (%) - Hybrid System Connected to Grid
Loading
(MW)
Grid
Connected
Islanding L-G Fault L-L Fault
Non -
Linear Load
Switch
0.875 1.2665 11.1161 3.5811 6.0342 62.3332
1 1.1347 8.0497 3.5966 6.0166 62.2358
1.3 0.953 3.3066 3.7453 5.9751 61.8576
1.75 0.7766 4.1653 3.9311 5.8795 61.3211
● Rated Power - 1.75 kW ( 0.25 MW + 1.5 MW )
● Loading varies from 50 % to 100 % of rated power

21. Inferences
THD can detect Islanding as well as Power Quality Issues
THD on Islanding event changes abruptly with load change
THD fails to maintain a threshold for Islanding
THD fails to differentiate Islanding and Power Quality Issues

22. Introduction to Wavelet Transform

23. Introduction to Wavelet Transform
Fourier Transform
Apt for Decomposition of Stationary
Signals
Basis function is Sine wave
Signal is represented as translated
and dilated versions of Sine
Wave
Frequency domain information is
available
Wavelet Transform
Apt for Decomposition of Non -
Stationary Signals
Basis function is Wavelet
Signal is represented as translated
and dilated versions of
Wavelet
Frequency and Time domain
information is available

24. Advanced Multirate Signal Processing
Non Stationary Signal
High Frequency Parts
Low Frequency Parts
High Frequency parts
Quick parts
Need more samples to detect them
Low Frequency parts

25. Discrete Wavelet Transform
v - Input Signal
Ψ - Mother Wavelet
- Translation Constant / Position
- Dilation Constant / Scale
Dyadic scale have =

26. Discrete Wavelet Transform - Working
● x[n] → Signal
● h[n] → HPF
● g[n] → LPF
Coefficients
Filters
Down - Sampling

27. Frequency Division
● Let sample frequency of signal = Fs
■ Eg. Fs = 1 kHz
Wavelet Level Frequency Band (Hz)
Level 1 500 - 1000
Level 2 250 - 500
Level 3 125 - 250
Level 4 62.5 - 125

28. Mother Wavelet Selection
Reconstruction capability
Empirically find if the input signal can be reconstructed by the wavelet perfectly
Similarity of Wavelet and Input Signal

29. Daubechies 4 Mother Wavelet
● Disturbance in power system features exhibit sharp changes
● Daubechies mother wavelet with low order
○ Which have an angular shape
○ Ideal to analyse sharp changes
■ For smooth features , a higher order is preferred.

30. Daubechies 4 Mother Wavelet

31. Statistical Indices Of Wavelet Transform
At a particular frequency band / level
Standard Deviation
Power of signal when its mean =0
L2 Norm ( Energy )

32. Wavelet Transform Based Approach
Extract Neg.
Sequence Voltage
from PCC
Perform Wavelet
Transform
Find SD and Energy
values of appropriate
levels
Fix the threshold
GRID CONNECTED MODE - UNDISTURBED SYSTEM

33. Wavelet Transform Based Approach
Extract Neg.
Sequence Voltage
from PCC
Perform Wavelet
Transform
Find SD and Energy
values of appropriate
levels
Detect Disturbance
by Comparing With
Threshold
SYSTEM UNDER DISTURBANCE

34. Frequency Levels Under Consideration
Negative Sequence Voltage taken from PCC
Sample Frequency taken - 1 kHz , Fundamental Frequency - 60 Hz
Wavelet Level Frequency Band (Hz)
Level 1 500 - 1000
Level 2 250 - 500
Level 3 125 - 250
Level 4 62.5 - 125
Level 5 31.25 - 62.5

35. SD3 - Hybrid System Connected to Grid
● Standard Deviation at level n - SDn
Loading
(MW)
Grid
Connected
Islanding L-G Fault L-L Fault
Non -
Linear Load
Switch
0.875 0.00011473 0.03401753 0.00494453 0.00911728 0.01280362
1 0.00011594 0.03204665 0.00494597 0.00911516 0.01280433
1.3 0.00010763 0.02510713 0.00495142 0.00911595 0.01279189
1.75 0.00012745 0.00583276 0.00495896 0.00912749 0.01277151

36. SD4 - Hybrid System Connected to Grid
● Standard Deviation at level n - SDn
Loading
(MW)
Grid
Connected
Islanding L-G Fault L-L Fault
Non -
Linear Load
Switch
0.875 0.00014444 0.03705120 0.01195818 0.02928778 0.02509573
1 0.00015432 0.04406543 0.01196308 0.02921008 0.02509539
1.3 0.00017652 0.02957596 0.01204052 0.02910395 0.02506549
1.75 0.00022712 0.00608325 0.01213379 0.02892913 0.02500862

37. E3 - Hybrid System Connected to Grid
● Energy at level n - En
Loading
(MW)
Grid
Connected
Islanding L-G Fault L-L Fault
Non -
Linear Load
Switch
0.875 0.00130841 0.39045041 0.05639256 0.10401861 0.14671166
1 0.00132418 0.36693973 0.05641079 0.10399256 0.14673016
1.3 0.00123171 0.29207070 0.05647472 0.10399785 0.14659020
1.75 0.00145480 0.06664993 0.05656272 0.10413084 0.14638457

38. E4 - Hybrid System Connected to Grid
● Energy at level n - En
Loading
(MW)
Grid
Connected
Islanding L-G Fault L-L Fault
Non -
Linear Load
Switch
0.875 0.00119563 0.30587180 0.09861007 0.24154712 0.20694654
1 0.00128137 0.36386052 0.09865040 0.24090679 0.20694387
1.3 0.00146062 0.24395503 0.09928896 0.24003145 0.20669716
1.75 0.00187444 0.05016677 0.10005795 0.23858983 0.20622875

39. Inferences
Change in SD and Energy are not abrupt, unlike THD
Thus it's easy to fix a threshold with SD and Energy
SD and Energy can classify the events
Between Islanding and Power Quality Issues

40. Machine Learning
Subfield of Computer Science
Study of pattern recognition
Learns from examples
Training → Learning → Prediction
Category

41. Applications
Voice assistants like Google Voice Search , Cortana and Siri
Product Recommendations on E- Commerce sites
Spam mail classifier
According to Gartner’s Hype Cycle 2015
Most promising technology of future

42. Artificial Neural Networks
Supervised Learning Technique
Learn from labelled examples
Classifier
Inspired by the Human Neural System

43. Input Vector ( Feature Vector )
● Loading of DG (MW)
● Standard Deviation at Level 4
● Standard Deviation at Level 3
● Energy at level 4
● Energy at level 3

44. Labels
Event Label
Grid Connected Mode 1
Islanding 2
L-G Fault 3
L-L Fault 4
Non Linear Load Switch 5

45. ANN - Specifications
No. of layers - 3
Input layer size - 5
No. of features taken
Hidden layer size - 40
Optimized by trial and error method

46. Training Set - Grid Connected Mode
Feature Vector (X)
Label (Y)
Loading
(MW)
SD3 SD4 E3 E4
0.875 0.00011473 0.00014444 0.00130841 0.00119563 1
1.015 0.00011594 0.00015432 0.00132418 0.00128137 1
1.295 0.00010763 0.00017652 0.00123171 0.00146062 1
1.505 0.00011485 0.00019470 0.00131076 0.00160678 1
1.75 0.00012745 0.00022712 0.00145480 0.00187444 1

47. Training Set - Islanding Mode
Feature Vector (X)
Label (Y)
Loading
(MW)
SD3 SD4 E3 E4
0.875 0.03401753 0.03705120 0.39045041 0.30587180 2
1.015 0.03204665 0.04406543 0.36693973 0.36386052 2
1.295 0.02510713 0.02957596 0.29207070 0.24395503 2
1.505 0.00600586 0.00491463 0.06873916 0.04055274 2
1.75 0.00583276 0.00608325 0.06664993 0.05016677 2

48. Training Set - L-G Fault
Feature Vector (X)
Label (Y)
Loading
(MW)
SD3 SD4 E3 E4
0.875 0.00494453 0.01195818 0.05639256 0.09861007 3
1.015 0.00494597 0.01196308 0.05641079 0.09865040 3
1.295 0.00495142 0.01204052 0.05647472 0.09928896 3
1.505 0.00496716 0.01207443 0.05665421 0.09956846 3
1.75 0.00495896 0.01213379 0.05656272 0.10005795 3

49. Training Set - L-L Fault
Feature Vector (X)
Label (Y)
Loading
(MW)
SD3 SD4 E3 E4
0.875 0.00911728 0.02928778 0.10401861 0.24154712 4
1.015 0.00911516 0.02921008 0.10399256 0.24090679 4
1.295 0.00911595 0.02910395 0.10399785 0.24003145 4
1.505 0.00911367 0.02904593 0.10397254 0.23955288 4
1.75 0.00912749 0.02892913 0.10413084 0.23858983 4

50. Training Set - Non Linear Load Switch
Feature Vector (X)
Label (Y)
Loading
(MW)
SD3 SD4 E3 E4
0.875 0.01280362 0.02509573 0.14671166 0.20694654 5
1.015 0.01280433 0.02509539 0.14673016 0.20694387 5
1.295 0.01279189 0.02506549 0.14659020 0.20669716 5
1.505 0.01278282 0.02503957 0.14650191 0.20648373 5
1.75 0.01277151 0.02500862 0.14638457 0.20622875 5

51. Data Set
No. of examples for each event - 26
No. of events = 5
Total Examples = 26 x 5 = 130
Splitting percentage of data - 70 %
Training Set - 91

52. Inferences
Verified using 3-fold cross validation
ANN can classify and identify the events with excellent accuracy
Prediction accuracy is > 95 %

53. Final Conclusions
WT Indices provide a better prospect on detecting and classifying Islanding
and PQ disturbances in a Hybrid DG System
THD fails to perform accurate classification
Implementation using Machine Learning Classifier
High prediction accuracy

54. Future Scope
Detection of Voltage Swell events are not well performed by WT
Load Switching
Capacitor Bank switching
Performance of WT at noisy environments
Feature Vector Selection

55. Implementation
Implementation as a Web Service
Using Microsoft Azure Machine Learning Suite
API Key -
zogVJpUKUmTEEh3auMuVYZ1q1tBKBWA+tHeZWIyzuG2sYUfxHKJ7F4IpXhGALe5IiVpunz
MHjDF8i0gTImfqfA==
URL -
https://ussouthcentral.services.azureml.net/workspaces/8e3a01b9b5a94cd8be8f69c85b
fd1215/services/923ca965d52a46788d91c32cd9cdc9fd/execute?api-
version=2.0&details=true

56. Web Service- Microsoft Azure ML Suite

57. GUI Using Python

58. GUI Using Python

59. Reference
[1] El-Saadany, E. F., H. H. Zeineldin, and A. H. Al-Badi. "Distributed generation: benefits and challenges." International Conference on Communication, Computer & Power. Vol.
115119. 2007.
[2] Algarni, Ayed. "Operational and planning aspects of distribution systems in deregulated electricity markets." (2009).
[3] Zeineldin, H. H., and M. M. A. Salama. "Islanding detection of inverter-based distributed generation." IEE Proceedings-Generation, Transmission and Distribution 153.6
(2006): 644-652..
[4] Eltawil, Mohamed A., and Zhengming Zhao. "Grid-connected photovoltaic power systems: Technical and potential problems—A review." Renewable and Sustainable Energy
Reviews 14.1 (2010): 112-129.
[5] United States of America. Congress of the U.S., Congressional Budget Office. Prospects for distributed electricity generation. September, 2003.
[6] Tailor, J. K., and A. H. Osman. "Restoration of fuse-recloser coordination in distribution system with high DG penetration." Power and Energy Society General Meeting-
Conversion and Delivery of Electrical Energy in the 21st Century, 2008 IEEE. IEEE, 2008.
[7] Xu, Wilsun, Konrad Mauch, and Sylvain Martel. "An assessment of distributed generation islanding detection methods and issues for Canada."CANMET Energy Technology
Centre-Varennes, Natural Resources Canada, QC-Canada, Tech. Rep. CETC-Varennes 74 (2004).
[8] Menon, Vivek, and M. Hashem Nehrir. "A hybrid islanding detection technique using voltage unbalance and frequency set point." IEEE Transactions on Power Systems 22.1
(2007): 442-448.
[9] PVPS, IEA. "Evaluation of islanding detection methods for photovoltaic utility-interactive power systems." Report IEA PVPS T5-09 (2002).
[10] Ropp, M. E., et al. "Using power line carrier communications to prevent islanding [of PV power systems]." Photovoltaic Specialists Conference, 2000. Conference Record of the Twenty-Eighth
IEEE. IEEE, 2000.

60. Reference
[11] Xu, Wilsun, et al. "A power line signaling based technique for anti-islanding protection of distributed generators—part i: scheme and analysis." IEEE Transactions on Power
Delivery 22.3 (2007): 1758-1766.
[12 Xu, Wilsun, et al. "A power line signaling based technique for anti-islanding protection of distributed generators—part i: scheme and analysis." IEEE Transactions on Power
Delivery 22.3 (2007): 1758-1766.
[13] Redfern, M. A., O. Usta, and G. Fielding. "Protection against loss of utility grid supply for a dispersed storage and generation unit." IEEE Transactions on Power Delivery 8.3
(1993): 948-954.
[15] Wrinch, Michael C. Negative sequence impedance measurement for distributed generator islanding detection. Diss. University of British Columbia, 2008.
[16] Yin, Jun, Liuchen Chang, and Chris Diduch. "Recent developments in islanding detection for distributed power generation." Power Engineering, 2004. LESCOPE-04. 2004
Large Engineering systems Conference on. IEEE, 2004.
[17] Smith, G. A., P. A. Onions, and D. G. Infield. "Predicting islanding operation of grid connected PV inverters." IEE Proceedings-Electric Power Applications 147.1 (2000): 1-6.
[18] Ropp, M. E., Miroslav Begovic, and A. Rohatgi. "Analysis and performance assessment of the active frequency drift method of islanding prevention."IEEE Transactions on
Energy Conversion 14.3 (1999): 810-816.
[19] Hung, Guo-Kiang, Chih-Chang Chang, and Chern-Lin Chen. "Automatic phase-shift method for islanding detection of grid-connected photovoltaic inverters." IEEE
Transactions on energy conversion 18.1 (2003): 169-173.
[20] Hernandez-Gonzalez, Guillermo, and Reza Iravani. "Current injection for active islanding detection of electronically-interfaced distributed resources."IEEE Transactions on
power delivery 21.3 (2006): 1698-1705.

61. Publications [1] Sleeba Paul Puthenpurakel, Subadhra P.R.,
“Islanding Detection in Grid-Connected 100 KW
Photovoltaic System Using Wavelet Transform”,
International Conference on Emerging Trends in
Smart Grid Technology - INCETS'16, IJIRSET Volume
5, Special Issue 5, April 2016
[2] Sleeba Paul Puthenpurakel, Subadhra P.R.,
“Identification and Classification of Microgrid
Disturbances in a Hybrid Distributed Generation
System Using Wavelet Transform”, International
Conference on Next Generation Intelligent Systems”

62. Thank you all very much !