Fault type and location detection in distribution networks with improved random vector functional link networks

Fault detection is an important subject in increasingly complex electrical transmission and distribution systems. In the present dissertation study, two methods based on random vector functional link networks (RVFLN), Cauchy weighting M-estimation RVFLN (Cauchy-M-RVFLN) and online robust regularized RVFLN (ORR-RVFLN) methods, are proposed for the detection of short circuit faults and fault locations, which are a major problem of transmission and distribution systems. The fault detection performances of the proposed methods were analyzed and compared with those of conventional machine learning methods and with each other. The performances of methods were analyzed with data obtained from simulations on a real-time digital simulator (RTDS). As a simulation model, IEEE 33 and IEEE 39-bus systems developed on RSCAD, a real-time simulation software package, were used. Different from studies in the literature, this dissertation study introduces six new feature vectors to be used in the detection of faults and fault locations. It was detected that these newly constructed feature vectors contributed to increasing the detection accuracy of high-impedance short circuit faults by about 10%. In addition, the performances of methods were tested by adding noise to data in order to achieve similarity to real systems. The proposed methods can detect faults with current and voltage data retrieved from only eight buses of the transmission and distribution system. The performances of methods were analyzed with data obtained for a large number of conditions including four different fault impedance values for different short-circuit faults which were 0, 30, 50 and 100 ohm and seven different distances from the bus which were 0-50, 50-100, 100-150, 150-200, 250-300, 350-400 and >400 m. Results showed that the proposed ORR-RVFLN and Cauchy-M-RVFLN methods detected the faults with 100% and 80% accuracies depending on the fault type, fault impedances and noise values. The proposed methods were able to perform fault detection with higher accuracy than RVFLN, convolutional neural networks (CNN), long short-term memory (LSTM), support vector machine (SVM) and extreme learning machine (ELM) methods. Results regarding the detection of fault locations showed that the number of erroneous fault location detections increased with higher fault impedance and decreased with longer distance from the bus. It was observed that the most erroneous detections of fault locations occurred in 0-50 m range which was the shortest distance from the bus. Study results demonstrated that the method detected the fault locations farther than 400 m with an accuracy of over 90%. When the proposed methods were compared with each other, the ORR-RVFLN method had a better fault detection performance than the Cauchy-M-RVFLN method by 3% and 8%. In the light of all these results, we can conclude that the proposed ORR-RVFLN method is the most successful method for the detection of faults and fault locations.