Comparative assessment of fault detection methods in transmission line during power swing: analyzing single and hybrid networks for the southern Kerala grid

The blocking function of power oscillation is a highly effective feature found in distance relays. Its purpose is to prevent unintentional tripping of transmission lines. It is crucial to accurately differentiate between faults and power oscillations to minimise the occurrence of widespread power outages and monetary losses. Therefore, the detection and classification of faults arising from power swings are critical challenges in ensuring the smooth operation and overall health of transmission lines, particularly in extra-high-voltage (EHV) and ultra-high-voltage (UHV) power systems. Modern methods rely on machine learning methods which are in their infancy. The paper proposes to identify various fault conditions in transmission lines during power swing by using a hybrid Neural Network architecture. In this research, different designs such as 2-D convolution neural network (2D-CNN), 1-D convolution neural network (1D-CNN), long short-term memory (LSTM), and CNN-LSTM hybrid networks are investigated to understand the effect of various deep learning approaches. Along with the southern Kerala grid, the standard 9-bus system is selected as a test case to examine various fault circumstances of the system during power swing. Multiple tests were conducted to determine the optimal deep learning architecture, including model parameters and configurations, for accurate fault detection. For the 9 bus system, the 1D CNN network performs better with an accuracy of 98.70%, and for the Kerala grid, both networks are competitive, but the CNN-LSTM Hybrid method slightly outperforms the 1D CNN with an accuracy of 92.40%.