An Overview of Transmission Line Protection by Artificial Neural Network: Fault Detection, Fault Classification, Fault Location, and Fault Direction Discrimination
Table 4
Comparative study of ANN based “faulty phase selection” schemes.
Two-level: (i) ANN-1 for fault type classification and ANN-2 for faulty phase selection for each fault type
1/4 cycle
(i) Feed-forward with 1 hidden layer. 1stlevel: 60 inputs with 90 neurons in the hidden layer and 4 outputs. 2nd level: 60 inputs with 30 neurons in the hidden layer and 3 outputs 1st level: training data extracted from the fault at 40% and 60% distance from the relay, while testing data was from the same fault at 20% 2nd level: training data extracted from the fault at 20% distance from the relay, while test data at 40% location
Authors used single circuit 500 kV study system and 2-level hierarchical neural network for higher learning ability and accuracy
(i) ANN architecture: (18-12-3) (ii) Hyperbolic tangent function (iii) 6 frequency ranges were chosen for each phase and then converted into six features (iv) Frequency ranges: (1) below 16.6 kHz; (2) over the range 16.7–31.3 kHz; (3) over the range 31.4–46.91 rHz; (4) over the range 50–65.6 kHz; (5) over the range 66.7–81.3 kHz; (6) over the range 81.4–100 kHz Sampling frequency: 200 kHz
Advantages: Information used within the phase selector was based on the fault generated high frequency noise Unlike conventional techniques, this information was not influenced by different systems and fault conditions However it was mainly dependent on the behavior of the nonlinear fault arc
(i) 6 inputs and 4 outputs and 7 and 4 neurons in the hidden layers (ii) Trained with both BP and Marquardt-Levenberg (ML) algorithms (iii) Activation function: Tan-sigmoid function was used for hidden layer neurons and saturated linear function for the output layer
ANN approach could perform well even in the presence of substantial amount of fault resistance for the far end faults
ANN based accurate fault phase selector and distance locator
Within a quarter cycle
(i) Fundamental components of current and voltage signals as input (ii) Three-phase current input signals were processed by simple 2nd-order low-pass Butterworth filter with cut-off frequency of 400 Hz. (iii) Hyperbolic tangent function in the hidden layer and purelin in output layer (iv) Bayesian regularization BP training algorithms
For most of the faults on the line, the location module is able to respond with an error less than 0.4 percent
SVM based fault phase selection and ground detection for fault type classification
10 ms (half-cycle)
(i) SVM1 for fault phase selection: Postfault current and voltage samples for one-fourth cycle (five samples) as input (ii) SVM2 for ground detection: zero sequence components of fundamental, third and fifth harmonic components of the postfault current signal (iii) Sampling frequency: 1.0 KHz Data collection: PCL-208 data acquisition card, which uses 12-bit successive approximation technique for A/D conversion; installed on a PC (P-4) with a driver software routine written in C++ having six I/O channels with input voltage range of +5 V
(i) The test results are compared with those of the radial basis function network (RBF) and were found to be superior with respect to efficiency and speed (ii) The classification test results from SVMs are accurate for simulation model and experimental setup and thus provide fast and robust protection scheme for distance relaying in transmission line
Combination of wavelet transform and neural network trained with Levenberg-Marquardt (LM) algorithm
—
(i) Sums of absolute values of 6th level detail coefficients (Db8) of line currents as inputs (ii) ANN architecture (8-15-7) with set mse goal of (iii) Hyperbolic tangent-hidden layers and pure linear-output layer (iv) Type of fault: LG, LL, LLG, LLL, and cross-country faults; fault location (km): 10 to 90 in steps on 10 km; fault inceptionangle: 0° to 315° in step of 45° (v) Fault resistance (Ω): 0 to 200 in steps of 25 Ω Sampling frequency: 12.77 KHz
Their proposed phase selector scheme can correctly identify faulted phase on the double circuit transmission line
ANN with BP and Marquardt training (Trainlm) algorithm
Half-cycle
Total 1400 training samples and 200 testing samples (A) Fault classification: 3 input, 3 output, and 8 nodes in hidden layer (B) Fault location: (i) Input layer: first 5 rows of the output of the S-transform with the line model current signals (ii) Hidden layers: 2 with 16 nodes each (iii) Activation function: tansig function for hidden layer and logsig function for output layer
Both faulty phases and healthy lines produced high frequency components because of mutual coupling, so here, S-transform energy of transient current was used to select faulty phase; ANN nonlinear fitting function was used to locate fault distance based on S-transform extracted transient energy
Multilayer feed-forward network and wavelet transform
1.2 ms
DB4 mother wavelet Training patterns: (i) Input layer: 30 inputs (10 level decomposition of 3 phases current components) (ii) Hidden layer: 20 nodes (iii) 4 outputs (a, b, c, and g) (iv) Sampling frequency: 16.7 kHz (v) Fault resistance (Ω): 2 Ω (vi) Busbars capacitor: 0.1 F (vii) Fault inception angle: 0° and 90°, / ratio = 100
Effectively classified the faulty phase(s) and healthy phase(s) just requiring 20-sample length window data (1.2 ms) and real-time implementation can be possible
Fault classification and fault location based on Back propagation algorithm (BPN), radial basis function (RBF) network and cascaded correlation feed-forward network (CFBPN)
—
(i) Sequence components of the fault currents of both sending end and receiving end as input (ii) Input samples of 1000 × 6 (iii) Fault type: LG, LLG, and LLLG (iv) FFBPN architecture (1-2-1)
Among all the ANN modules, results of RBF network were found to be better than the other two networks in terms of accuracy
