부산대학교 도서관

Adaptive protection schemes (APS) have gained prominence in maintaining the integrity of overcurrent relay (OCR) settings in reconfigurable networks. While many APSs rely on supervisory control and data acquisition systems, they are very expensive and expose the system to vulnerabilities arising from communication failures. Recent studies have proposed communication-less APSs to address this issue by relying on data-mining algorithms equipped with real-time fault voltage-current information. However, the OCR settings are computed and updated as the fault occurs, inevitably causing prolonged OCR tripping in these schemes. This contradicts with the APS’ original purpose of minimizing OCR operation time and consequent equipment damage. Thus, a load flow-based APS that addresses this flaw is proposed to achieve primary-backup OCR coordination in a highly reconfigurable system. Network topologies are first categorized into OCR setting groups via clustering analysis. A nonparametric probability model is developed to evaluate the probability of network topologies at a measured load flow. Then, a machine learning model deployed in a local controller selects the correct setting groups based on the calculated probabilities. The proposed APS achieves high accuracies and low OCR operating times in the IEEE 33-bus test distribution system under varying load conditions and network topologies.