|
| Author and year (reference) | Method used | Response time | ANN features | Accuracy |
|
| Dalstein and Kulicke, 1995 [34] | ANN architecture and digital signal processing.
Simulation program
NETOMAC | (i) Average classification time <6 ms
(ii) Arcing fault detection time: 25 ms to 70 ms | Training patterns: 2268 fault cases
(i) Two hidden layers with (30-20-15-11)
(ii) Back propagation training algorithm
Test cases: 240 | — |
|
|
Kezunovic and Rikalo, 1996 [35] | Combined supervised and unsupervised
neural network with ISODATA clustering algorithm | Fault detection logic: 0.2 ms and fault classification logic: 15 ms | (i) Training patterns: 1189
(ii) 2 kHz sampling rate
(iii) ISODATA clustering algorithm based unsupervised neural network
(iv) Test cases: 1188 | 67.93% to 94.36% fault classification rate |
|
|
Kezunović et al., 1995 [36] | Unsupervised neural network algorithm MINNET using DFR assistance | Fault detection time: 1 cycle | (i) Training patterns: 1200 faults cases
(ii) Two hidden layers with (30-20-15-11)
(iii) ISODATA clustering algorithm
(iv) Test cases: 295 | — |
|
| Vazquez et al., 1996 [37] | Feed-forward neural network (FFNN) | Fault detection time: 1/2 cycle | (i) Training patterns: 976–1464 faults cases
(ii) 960 Hz sampling rate
(iii) Single hidden layer network (5-20-1) or (10-20-1) | — |
|
| Vazquez et al., 1996 [38] | Feed-forward neural network (FFNN) | 1/4–1/2 cycle | (i) Training patterns: 976–1464 fault cases
(ii) 960 Hz sampling rate
(iii) Single hidden layer network (5-20-1) or (10-20-1) | — |
|
| Chowdhury and Wang, 1996 [39] | Kohonen neural network | 1 cycle | (i) Training patterns: 414, testing patterns: 206
(ii) Sampling frequency: 6 kHz
(iii) Using the fundamental components of currents and voltages
(iv) 2-dimensional Kohonen map consisting of 16 neurons | 100% accuracy |
|
| Chowdhury and Aravena, 1998 [40] | Kohonen network and multiresolution wavelet filter banks | Not mentioned | (i) Unsupervised Kohonen neural network
(ii) Daubechies’ wavelet of order ten for preprocessing of the voltage and current signals | — |
|
| Wang and Keerthipala, 1998 [41] | Fuzzyneuro approach to fault classification | <10 ms | Inputs: symmetrical components in combination with three line currents
(i) Combined fuzzy logic and neural network approach | — |
|
| Keerthipala et al., 2000 [42] | Fuzzyneuro approach to fault classification | <1 cycle | Three line current and symmetrical components of currents used as input to fuzzyneuro based protective relay using a real-time digital simulator | — |
|
| Vasilic and Kezunovic, 2005 [43] | Fuzzy ART neural network algorithm: fuzzy K-nearest neighbor classifier | 1/2 cycle–1 cycle | (i) Sampling frequency: 1.92 kHz and 3.84 kHz
(ii) Training patterns: 3315
(iii) Testing patterns: 5000 | — |
|
| Zhang and Kexunovic, 2005 [44] | Fuzzy ART neural networks | 1 cycle | (i) Sampling frequency: 1.92 Hz
(ii) NN1 for fault detection, NN2 for fault classification, and NN3 for ground identification
(iii) Training patterns: NN1-9564, NN2-9240, NN3-1152.
(iv) Testing patterns: 6000 | — |
|
|
Silva et al., 2006 [45] | Fault detection and classification using oscillographic data by ANN and wavelet transform | — | (i) Multilayer perceptron (MLP) Daubechies 4 (db4)-ANN architecture (40-30-4)
(ii) Training algorithm: RPROP with 426 iterations and rms error is 0.02
(iii) Sampling frequency: 1200 Hz
(iv) Fault type: LG, LL, LLG, and LLL
Test cases: 720 | (i) Fault classification accuracy: 99.83% |
|
| He et al., 2006 [46] | Wavelet entropy measure | 1 cycle | Wavelet entropy distributing along time to detect the transient fault by comparing with threshold or using ANN classifier | — |
|
|
Martín et al., 2008 [47] | Combined wavelet and ANN | 1/4 cycle | (i) Sampling frequency: 1.92 Hz
(ii) Current signals preprocessed using Mallat algorithm for DWT (Daubechies 6)
(iii) ANN based fault classification (24-24-7) | — |
|
| Kawady et al., 2008 [48] | A Gabor transform-ANN based fault detector | 1 cycle | (i) Sampling frequency: 6.4 kHz, GT as a feature extractor and artificial neural networks for pattern recognition and classification (64-10-2)
(ii) Test cases: 300 | — |
|
| Mahmood et al., 2008 [49] | Wavelet multiresolution analysis and perceptron neural networks | — | (i) Sampling frequency: 20 kHz
(ii) Daubechies eight (db 8): 5th level detail coefficient of current signals
(iii) Perceptron neural network (64-3) with hard limit and perceptron learning rule
(iv) Identify lightning stroke, switching surge, or fault | — |
|
| Geethanjali and Priya, 2009 [50] | Combined wavelet transforms and neural network | — | (i) Daubechies two (db 2): 5th level detail coefficient of current signals, three-layered ANN (16-33-3)
(ii) Training: 160 data, test: 20 data | 93.89% accuracy |
|
| Pothisarn and Ngaopitakkul, 2009 [51] | DWT and back-propagation neural networks | 1/4 cycle for fault classification | (i) Sampling rate is 200 kHz, db4 with 1st level detail coefficient and zero sequence components of voltage and current signals (8 inputs)
(ii) Training: 760 data, test: 360 data
(iii) Two hidden layers ANN (8-8-11-4) | 97.22% accuracy |
|
| Abdollahi and Seyedtabaii, 2010 [52] | Comparison of Fourier and wavelet transform methods for fault classification | 3 cycles | Sampling frequency: 1 kHz
Energies of detailed DWT coefficients of 1st and 2nd frequency components of Iabc and Izero are summed together to form inputs
(i) Training: 150 data, test: 50 data
(ii) Feed-forward BPNN for fault classification | 98% accuracy |
|
| Jain et al., 2010 [53] | Artificial neural network for intercircuit and cross-country fault | <1 cycle | Sampling frequency is 1 kHz, fundamental components of voltage and current signals as input to feed forward neural network (FFNN) with Levenberg-Marquardt algorithm (9-50-7)
(i) Training and testing: 6000 patterns | — |
|
| Jain et al., 2008 [54] | ANN based fault detector and classifier | <1 cycle | Sampling frequency is 1 kHz, superimposed, zero and negative sequence components of current signals as input to three layers FFNN trained with Trainlm algorithm (10-10-7), test cases: 240 | 100% accuracy |
|
| Jain et al., 2009 [55] | ANN based fault
classifier for SLG faults | <1 cycle | Sampling frequency is 1 kHz, fundamental components of voltage and current as input
(i) Training: 1840 and testing: 1800 patterns ANN based fault detector and classifier with single hidden layer (9-30-7) | 100% accuracy |
|
| Jain et al., 2010 [56] | ANN based fault classifier and locator | <1 cycle | Sampling frequency is 1 kHz, fundamental components of voltage and current as input
ANN-FC-Kohonen self-organising map
ANN-FL-FFNN (Bayesian regularisation algorithm) | — |
|
| Yadav, 2012 [57] | Comparison of single and modular ANN based fault detector and classifier | <1 cycle | Sampling frequency is 1 kHz, fundamental components of three voltage and six currents of double circuit line as input to ANN
Training patterns: 8800, testing patterns: 2400
Single ANN (9-50-7), modular fault typewise: ANN LG (9-30-7), LL (9-30-7), LLG (9-8-7), and LLL 9-20-7
Trainlm algorithm | (i) 98% accuracy
(ii) Detects/classifies intercircuit, cross-country and evolving faults |
|
|
Jain, 2013 [58] | ANN based fault detection for transmission lines | <1/4 cycle | (i) Sampling frequency is 1 kHz, fundamental components of three voltages and three currents as input to ANN. Single ANN (6-10-10-1), Trainlm algorithm
(ii) Training patterns: 29123, testing patterns: 26912 | (i) 100% accuracy |
|
| Chen and Aggarwal, 2012 [59] | Wavelet transform and artificial intelligence | 1-cycle data (20 samples) | (i) Sampling rate 16 kHz, spectral energy details of 10-level DB wavelet coefficients of 3-phase current as input (30) to ANN (30-20-4) | — |
|
|
Ben Hessine et al., 2014 [60] | Artificial neural networks | 1 cycle | (i) Sampling frequency: 1 kHz, fundamental and zero sequence components of three voltages and three currents as input to ANN FD (8-16-1),
(ii) 4 fault classifiers: ANN-R (8-5-1), ANN-S (8-5-1), ANN-T (8-5-1), and ANN-G (8-6-1) | — |
|
|
He et al., 2014 [61] | A rough membership neural network approach for fault classification | | (i) Sampling rate is 50 kHz, wavelet energy of three-phase currents and zero sequence current as input to RMNN (13-14-1) 10 separate RMNNs for 10 types of fault
Test cases: 60 | Average success classification rate of 99.4% |
|
| Koley et al., 2011 [62] | ANN for detection and classification of faults on six-phase transmission line | <1 cycle | Sampling by 1.2 kHz, fundamental components of six-phase voltages and currents
(i) Training patterns: 1460
(ii) ANN (Trainlm algorithm) (12-40-7) | — |
|
| Koley et al., 2012 [63] | ANN for six-phase to ground fault detection and classification of transmission line | <1 cycle | (i) Sampling by 1.2 kHz, fundamental components of six-phase currents, Training patterns: 370
(ii) ANN (trained with trainlm) (6-3-7) | — |
|
| Koley et al., 2012 [64] | ANN for phase to phase fault detection and classification of six-phase transmission line | <1 cycle | (i) Sampling by 1.2 kHz, fundamental components of six-phase voltages and currents
(ii) Training patterns: 4850
(iii) ANN model (12-30-6), Trainlm algorithm | — |
|
|
Koley et al., 2014 [65] | ANN based protection scheme for shunt faults in six-phase transmission line | <1 cycle | Sampling by 1.2 kHz, fundamental components of six-phase voltages and currents
(i) Total 22 modular ANN modules for fault detection/classification and distance location
(ii) Testing cases: 4930
(iii) Trainlm algorithm | 100% accuracy
Fault location error ±0.73% |
|
| Kumar et al., 2014 [66] | Haar wavelet and ANN based phase to phase fault classification in six-phase transmission line | — | Sampling by 1.2 kHz, standard deviation of approximated Haar wavelet coefficients of six-phase voltage and currents as input
(i) Training patterns: 1220
(ii) Testing patterns: 100
(iii) ANN model (12-5-6) trained with Trainlm | — |
|
