|
| Author and year (reference) | Method used | Response time | ANN features | Remark(s) |
|
| Sidhu et al. (1995) [87] | Multilayered feed-forward neural network (MLP) | 2.4 ms | (i) Three-layer MLP with sigmoid transfer function
(ii) Training dataset: 3240
(iii) Preprocessing: samples were processed by 4th-order low-pass antialiasing filters at 24 kHz and were resampled at 1.2 kHz with 100 Hz cut-off frequency
(iv) ANN based discriminator was implemented on a TMS320C30 based system | The direction determination was not affected by the type of fault, phases involved, power flow conditions, location of the fault, variation in source impedances, the presence of fault resistance, and missing data samples |
|
|
Sanaye-Pasand and Malik (1996, 1997) [88, 89] | Back propagation (BP) and Marquardt-Levenberg (ML) learning algorithm were compared and ML was chosen because of the reduced network error | — | (i) A bandpass 2nd-order Butterworth filter with 60 Hz passband, three-phase voltage and current sampled at 1.2 kHz (20 samples/cycle)
(ii) 30 inputs, two hidden layers (10, 5) of neurons and one output
(iii) A new smaller network with 20 inputs, two hidden layers with (10, 5) needs 30% less number of epochs to reduce the error to 5% of its initial value [88] and to 15% in [89] | (i) Authors found that the new 20-input network performed better than the earlier 30-input network
(ii) Here, real world fault data had been recorded by Alberta Power Ltd. on the 240 kV transmission systems [89] |
|
| Sanaye-Pasand and Malik (1998) [90] | Elman network for fault direction estimation | 0 ms–12 ms | (i) Elman network is a two-layer feed-forward network with the addition of a recurrent connection from the output of the hidden layer to its input
(ii) The network outputs which fall above 0.5 and below −0.5 are interpreted as forward and backward faults, respectively
(iii) 12 inputs, 12 hidden neurons, and one output neuron
(iv) For both hidden and output layers: tansig function | 40 different forward and backward faults at the relay location were applied to the system and the network’s performance was investigated; in all cases except one 3 phases to ground fault, the directional module performed correctly |
|
|
Wang et al. (1997) [91] | Three-layered multilayer feed-forward network with BP algorithm | — | (i) 500 KV, 300 KM transmission line tested under different operating and fault conditions
(ii) Sampled at 1.2 kHZ
(iii) Three-layered multilayer feed-forward network with BP algorithm (14 inputs and 5 output values) was used
(iv) Training patterns: 3172 | Directional comparison power line carrier protection based on the ANN was highlighted in this paper |
|
| Song et al. (1997) [92] | Combined genetic algorithm and ANN for fault direction of UPFC transmission line | — | (i) GANN: employs a feed-forward NN with GA training
(ii) Fault patterns: 6000
(iii) The NN is composed of 12 inputs (3-phase voltages and currents with 2 samples data window), 8 hidden neurons, and 4 outputs (roughly indicating fault position)
(iv) The population size (each weight contains the number of population) is varied from 20 to 100
(v) The parental bias parameter was set to 1.4
(vi) Mutation probability was set to 0.3 and the crossover probability was set to 0.8 | Disadvantages of BPNN over GANN:
BPNN needs larger training set covering data of various fault conditions (slow and time consuming); GA can be used for weight optimization
Disadvantages of GANN:
GANN training is also a time consuming process as there are a number of populations for each weight; but in this study, the GANN training is off-line, so time consumption does not matter as long as it can achieve better classification
Average misclassification rate: GANN: 2.35% BPNN: 3.70% |
|
|
Fernandez and Ghonaim (2002) [93] | Finite impulse response artificial neural network (FIRANN) for fault detection and direction estimation | 2.5 to 4.5 ms | (i) Only unfiltered voltage and current signals sampled at 2 kHz as input
(ii) Training patterns: 50000 fault patterns consist of a prefault cycle (40 samples) and 1.25 postfault cycles (50 samples)
(iii) Testing: 100,000
(iv) Temporal back propagation algorithm
(v) ANN architecture: (8-45-35-5)
(vi) Number of time-delay units:
(5, 2, 2); activation function: symmetric sigmoid | The relay is called FIRANN DSDST and is based on FIRANN type; the relay can detect the fault, determine the faulty phase, fault direction, and detect whether the fault is an undervoltage or undercurrent/overcurrent fault |
|
| Lahiri et al. (2005) [94] | Modular neural network approach | 3 samples | (i) Current samples of three phases and voltage samples as inputs (12) with 100 training patterns.
(ii) Six ANNs each with (10-3-1)
(iii) Implemented on a DSP TMS320F243 EVM-board with sampling rate of 1 kHz for 50 Hz system | The modular ANN concept reduces task-complexity and eliminates redundant inputs for fault classification |
|
Yadav and Thoke (2011)
[95] | ANN with Levenberg-Marquardt (LM) optimization learning algorithm | Less than 1.5 cycles | (i) Voltage and current available at only the local end of line
(ii) Training patterns: 1090
(iii) Testing patterns: 1090
(iv) For fault distance location task, 18 inputs and 8 and 7 neurons in hidden layer for FL1 and FL2, respectively, and 1 in the output layer were found to be suitable | (i) The proposed scheme allows increasing the reach setting up to 90% of the line length
(ii) It has the operating time of less than 1.5 cycles as it uses the one-cycle DFT
(iii) The technique does not require communication link to retrieve the remote end data |
|
| Yadav and Swetapadma (2014) [96] | ANN with Levenberg-Marquardt (LM) optimization learning algorithm | Fault detection, direction estimation, and fault classification take less than half-cycle time | (i) Fundamental component of current and voltage signals at one end of line as input
(ii) 3 ANNs for fault detection, classification, and direction estimation
(iii) ANN based fault detector (9-20-20-1)
(iv) Training and testing fault cases 40800 | (i) Main advantage of scheme is that reach setting of relay is up to 99%
(ii) Not affected by variation of parameters like fault type, fault resistance, fault location, fault inception angle, and so on
(iii) Provides primary and backup protection |
|
