An Overview of Transmission Line Protection by Artificial Neural Network: Fault Detection, Fault Classification, Fault Location, and Fault Direction Discrimination
Table 2
Comparative study of ANN based “fault detection and classification and location” schemes.
Multilayered backpropagation neural network for fault detection, classification, and location
(i) Average is 13 ms (ii) Classification module: 4 ms to 9 ms (iii) Location module: 8 ms–15 ms
(i) Multilayer perceptron with hyperbolic tangent activation function and supervised BP algorithm with Norm-Cum-Delta learning rule (ii) Sampling frequency: 1 kHz, test cases: 405 (iii) Fault detection module (ANN1 24-9-2) (iv) Fault classification module (ANN2 24-16-4) (v) Fault location module (ANN3 24-48-44-3, ANN4 24-44-40-3, ANN5 24-44-40-3, or ANN6 24-24-20-3)
The reach for protection zones 1, 2, and 3 is set at 95%, 130%, and 150% of the protected transmission lines, respectively. (i) Accuracy: 98% (ii) Error: 2%
Modular ANN approach for fault detection, classification, and location
ANN1: 2 ms detection time ANN2: 4 to 12 ms time to classify ANN3, ANN4, and ANN5 location time (8 to 15 ms) This scheme can operate in an average time of about 13 ms
(i) Modular ANN approach with hyperbolic tangent transfer function (ii) Learning rate: 0.01 to 0.4 (iii) Momentum: 0.001 to 0.2 intervals (iv) Sampling frequency: 1 kHz, test cases: 405 (v) Fault detection module: (ANN1 24-9-2) (vi) Fault classification module (ANN2 24-16-4) (vii) Fault location module: (ANN3 24-48-44-3, ANN4 24-42-40-3, or ANN5 24-24-20-3)
Fault classification accuracy is 99.44%. ANN relay estimated the expected response in approximately 98% of the 4,050 patterns tested. An extension of the relay primary protection zone to 95% of the line length was implemented
To detect fault using MRA wavelet transforms, three classifiers are used, namely, GRNN, PNN, and ANFIS The integral square error and multiple objective functions are used as a fitness function during the minimization operation
—
(i) Feed-forward NN: BFGS (quasi-Newton BP) is the training algorithm to classify the location of the fault (ii) GRNN: wavelet coefficient level 5 as input (iii) PNN: wavelet coefficient level 5 for length line of 5 percent increment as an input. (iv) ANFIS: 2 inputs and 1 output with 40 membership functions (v) Coefficients are range influence:0.05, accept ratio: 0.5, squash factor: 1.25, and reject ratio: 0.15
GRNN and PNN: 100% accuracy when used as fault classifier Fault location: PNN: 100% accuracy. If noise increased to 0.2% accuracy becomes 85% GRNN: 87% accuracy, but addition of 0.02% noise makes it 67.5% Feed-forward NN: 47.5% ANFIS: 82.5% correct classification
Radial basis function network (RBFN) with Gaussian transfer function was used for fault classification and location
—
(i) ANN based fault classification: only samples of three-phase currents as input (ii) Fault location: samples of both voltages and currents of the three phases as input (iii) Training for fault classifier: 120 data sets Fault location: ANN-I (for faults occurring beyond 50% of line) and ANN-II (for faults occurring within 50% of line)
Fault locator: error goal of 0.001 was considered The analysis result showed that the proposed system has good accuracy and validity
SARENEUR software tool was employed. An AMD Athlon 900 MHz computer with 128 Mb RAM was used to obtain the times. For each line, a total of 23028 cases were verified. These cases correspond to faults simulated for the three phases, in 101 positions with 76 different fault resistances. Several faults provided by the Spanish utility IBERDROLA S.A. were analyzed
—
Fault location: ANN has two hidden layers (8 to 9 neurons in 1st layer and 4 to 6 in the 2nd layer). There was no activation function in input layer, but LLP, LTP, TLP, or TTP activation functions were chosen in the output layer. The training time of ANN was always less than 3 min.
No classification error was found in single lines and the error was less than 1% in double circuit lines Mean error: Fault location: varies between 0.015% and 0.4% Fault resistance estimation: varies between 0.017% and 0.46%
Distributed and hierarchical NN (DHNN) system was implicated in this study, comprising IDNN and FLNN. IDNN (fault identification NN for detection and classification)
—
FLNN (fault location NN) for four fault classes (LG, LL, LLG, and LLL). (i) 7744 fault patterns. The output of fault location is processed through fuzzy technique which serves as control for accurate fault location. (ii) The location of FLNN has max = 0.754 kM and average absolute error mean = 0.2946 kM.
The study highlighted the utility of DHNN in identification and location of fault. It was evident from this study that location results were not influenced by fault sites, the intermediate resistances, the fault incidence angles, the opposite system impedance, and the phasor angles between EMF of the two systems
Back propagation algorithm and Levenberg-Marquardt algorithm
Faults are detected and classified within a quarter-cycle
(i) Only current signals measured at local end have been used to detect and classify the faults in the double circuit transmission line. (ii) Training patterns: (1) Fault type: A1G and A2G (2) Fault location, Lf (km): 0, 10, 20, 30, …, 80 and 90 km (3) = 0 and 90 deg (4) = 0 Ω, 50 Ω, and 100 Ω (iii) ANN architecture: 10-10-7 with mse of 5.22567
Performance of the protection technique has been illustrated with reference to only a single-phase-earth fault as this is the most frequently occurring fault (over 90% of all faults) in transmission networks
MRA wavelet transform (Daubechies 5) and probabilistic neural network (PNN). The model power system considered for the analysis is using Kundur’s four-machine two-area test system
—
PNN was used as the fault classifier (i) Fault type: one-phase fault. The designed algorithm was then run with the sigma equals to 0.01 and succeeded in obtaining 100% accuracy using both the training and test data.
Effect of noise addition on accuracy in PNN: 110 km line: 0.1% noise → 100% Acc 0.2% noise → 97.14% Acc 0.5% noise → 92.38% Acc 30 km line: 0.1% noise → 100% Acc 0.2% noise → 100% Acc 0.5% noise → 92% Acc
Radial basis function (RBF) based SVM and scaled conjugate gradient (SCALCG) used
—
It is a hybrid approach having two steps. Step 1: RBF based SVMestimates the initial distance of fault using the positive sequence voltages and currents of faulty phases. Step 2: Improving the final estimation of thisdistance using SCALCG based neural network with the high frequency range characteristics.
The maximum error of fault location was limited to 1.93 km in the worst case and 0.0001 km in the best case with the short duration of time in each 150 km line
(i) NeuroShell2 software was used to provide BP neural networks with structures as 6-5-5-3, 6-6-6-3, 6-7-6-3, and 6-5-4-3 (ii) Input layer is linear while at hidden layer and output layer is logistic function. (iii) BP network with two hidden layers
BP neural network architecture is an alternative method for fault detection, classification, and isolation/location in a transmission line system
A hybrid framework involving fault detection, classification, and location using SVMs and adaptive structural neural network (ASNN)
The detection of fault was performed in around 0.0005 s and one-cycle time period was needed to identify and locate the fault
(i) Fault samples: 240000 (ii) Positive, negative, and zero sequences as inputs (iii) SVMs and ASNNs (6 ASNNs each having 50 neurons). After fault detection, a multilevel wavelet transform was applied and features were obtained by PCA
Average detection accuracy of 99.9%, sensitivity and specificity for fault classification of 99.78% and 99.87%, respectively; and average fault location error of 0.47%
ANN for fault classification and fault distance location using Levenberg-Marquardt training algorithm
Fault is detected and classified within one cycle
(i) 220 KV Teed transmission circuit. Training cases (1) Fault type: ABG, BCG, and CAG (2) Fault location: in step of 10 km in each section (3) : 0° and 90° (4) : 0, 50, and 100 Ω (ii) Sampling frequency: 1 kHz-2nd-order low-pass Butterworth filter with cut-off frequency of 400 Hz (iii) mse goal reached at 1.093827
The proposed algorithm used the voltage and current signals of each section measured at one end of Teed circuit to detect and classify double line to ground faults (i) Automatic determination of faulted types and phases after one cycle from the inception of fault was achieved (ii) Algorithm eliminates the effect of varying fault location, fault inception angle, and fault resistance
An accurate fault classification and distance location algorithm for Teed transmission circuit based on ANN
The algorithm provides automatic determination of fault type, faulty phases, and fault distance location after one cycle from the inception of fault
(i) Levenberg-Marquardt training algorithm (ii) Mean square error goal reached 0.001 (1) Fault type: LG, LL, LLG, and LLLG (2) Fault location in step of 10 km in each section (3) : 0° and 90° (4) = 0 Ω, 50 Ω, and 100 Ω (iii) Three-layered ANN with 18-13-7 architecture
(i) The errors in locating the fault are in the range of −0.7% to +1.92%. (ii) The proposed scheme allows the protection engineers to increase the reach setting (i.e., a greater portion of line length can be protected)
ANN for fault distance location using Levenberg-Marquardt training algorithm
—
Levenberg-Marquardt (Trainlm) optimization technique for training of ANN based FL Training: 80% Validation: 10% Testing: 10% (24 data sets considered in testing)
Mean value of percentage error: fault location: 6.6%; fault resistance: 4.3% In most of the cases the error percentage to locate fault and to estimate resistance was less than 10 %
Combined wavelet transform and generalized neural network for fault location
—
(i) MRA based on DWT (Db4) for capturing the transient characteristics of the fault current signal (ii) : 10 to 1000 and = 1 to 10, : 36°, 54°, 90°, and 180° (iii) Sampling frequency: 100 kHz
Mean value of absolute relative error: Wavelet-GNN: around 2%, Wavelet-ANN: around 3% GNN model is more accurate than ANN
