Neural network pattern recognition schemes for identification and location of faults in thyristor controlled series-compensated (TCSC) HV power transmission lines
MetadataShow full item record
This thesis reports the development, implementation and evaluation of fault identification and location modules for protecting HV power lines with Thyristor Controlled Series-Capacitor (TCSC). The developed modules make use of the key capabilities of Artificial Neural Networks (ANN) and various attributes to approach their highly accurate decision. Multilayer Perceptron NN (MLPNN), used in this study, has the capability of learning and approximating any nonlinear function from a body of observations representing the problem at hand. Prior knowledge on the fault location problem is embedded in the training sets and used as constrains for developing the ANN. Two different novel techniques for addressing fault classification and location problems are considered. One of them makes use a single network with the optimal information embedded in one phase voltage and three phase currents. The second scheme is based on Modular neural network (MNN) where the main problem is divided into subtasks and each task is handled by its own individual NN. The first algorithm is characterized by two parallel networks: one dedicated for fault identification problem and the other one for fault location. The identification network utilizes a quarter cycle information in phase A voltage, three phase currents and neutral current as inputs. Once the fault is detected, the fault location network is triggered for locating the fault in transmission line with respect to the TCSC. The fault location network uses half cycle information of phase A voltage and three phase currents. This variable data window length makes the decision of both networks based only on the available local measurements independent of the thyristor firing delay angle, which has been considered as a key factor in the prior art. The proposed technique has been trained and tested through computer simulation studies for a typical two machine power system model implemented in EMTP-ATP. The transient effects of instrument transformers have been investigated. Simulation studies have also been considered for different operating conditions, including high fault resistance, fault location, compensation level and pre-fault power flow directions. This scheme has been found to be superior in terms of accuracy and speed over state-of-art commonly applied technique which employs all three phase voltages and currents as inputs to ANN. A different novel scheme based on modular NN has also been developed and investigated for fault classification and location problems. The fault classification task is divided into four separate subtasks, where the state of each phase and ground is determined by an individual neural network. The network for each phase is supplied by its respective voltage and current samples, whereas the decision of ground network is based only on the neutral current. The classifier outputs are post-processed by a logic circuit for triggering the proper NN in a fault location module. In this technique, the fault is located by three modular networks, one for each phase and fed by a recursive half cycle of related phase voltage and current signals. A physical, small-scale, two machine transmission system with TCSC in the center of transmission line has been designed and developed for evaluating and testing the protection schemes. The TCSC is modeled to work in four different operating modes. The modular NN approach has been verified through using LabVIEW software and NI-PCI6221 data acquisition card. Experimental results confirm the feasibility of the techniques for classifying and locating faults on power lines with TCSC through using only local measurements.