15-01-2013, 04:42 PM
Artificial Neural Network applications for Power System Protection
Artificial Neural Network.ppt (Size: 529.5 KB / Downloads: 21)
Transmission Line Protection:*
A multi-layer feed forward network with a 12-4-1 configuration was used in this design.* A recent design uses a Finite Impulse Response ANN (FIRANN) for detecting the onset of faults and determining the direction of the fault on high-voltage transmission lines.* Configuration of the proposed network is 45-35-5. Three of the five outputs of the network identified faults of the three phases (one for each phase); the fourth output determined the direction of the fault and the fifth output identified undercurrent/undervoltage conditions.* Processed sampled values of voltages and currents to determine if a fault is on the line side of the relay or is on the bus side of the relay.
Proposed Methodology:
Fully exploits the potential of ANNs and makes the whole pro cess simple.
* The processes assigned to the different layers of an ANN are segregated by dividing the network into sub-networks.
* Each sub-network is responsible for performing an assigned protection function.
* Helps in better understanding the internal structure of the ANN.
* Makes the pro cess of modifying the network simpler whenever required
Proposed Design:
* Instead of providing the conventional inputs (Va, Vb, Vc and Ia, Ib, Ic) inputs that assist in achieving the desired relay characteristics were used to train the network.
* A direct relationship between the inputs and the outputs expected from the networks while maintaining the integrity of operation around the boundary of the relay characteristics.
Modifications
One obtained output from each sub-network identified faults for the assigned(A, B or C fault).Sub-network was modified If an ANN did not work well during the testing stage.
b. Outputs from the three AND logic comparators were combined to detect
A-B, B-C and A-C faults.
c. Combining the outputs from all three networks detected three-phase
faults,an AND neuron was used by fixing specific weights of the neuron
instead of using neural layers as the next stage.