Monitoring Transient Stability in Electric Power Systems using a Hybrid Model of Convolutional Neural Network and Random Forest /

Abstract

With the rapid growth in power systems and extensive application of wide-area measurements, the security of the power system becomes a crucial aspect. Monitoring transient stability is required for reliable operation of the grid and instability prediction. The major idea of this research is to implement a fast and robust online transient stability assessment tool for the classification of the system operating states. Artificial Intelligence (AI) has drawn the attention of the power industry and is now being used to build a robust Transient Stability Assessment (TSA) system. Compared with traditional ANN techniques, this research work proposes an advanced hybrid TSA monitoring system which is a combination of a Convolutional Neural Network and Random Forest classifier. A Convolutional Neural Network (CNN) utilizes images for detecting the transient stability status and is hereby used as a feature extractor for the Random Forest (RF) classifier. This research determines that the combination of these two models can increase the efficiency of the system. A dataset is obtained by the Phasor Measurement Unit (PMU) measurements through time-domain simulations. Owing to the huge size and unapproachability of transmission systems, real-time testing can become problematic. Therefore, the simulations are performed on the standard IEEE-14 bus system and the IEEE-39 bus system as case studies to analyze the system for future practice in a lab and real-time grid. The efficiency of the system is tested under various contingency scenarios by implementing a number of faults in certain locations under varying loading conditions. A doubly-fed induction generator wind turbine is also installed to examine the intermittent effects on the power system stability. Moreover, a CNN-based fault classification system is also proposed for the identification of a fault that occurred in the system. The test results verified that the proposed hybrid model is accurate and more robust as the RF classifier is proved to have higher precision and required less time as compared to the CNN-based classifier. A combination of these techniques exhibited excellent results.