Development of Fault Location, Isolation and Service Restoration Algorithm by Considering Keypoint Equipment Indications at PLN UID West Kalimantan

Conventional Fault Location, Isolation, and Service Restoration (FLISR) algorithms predominantly assume that all switching devices operate under ideal conditions, neglecting unhealthy device anomalies such as mechanical actuator failures, low pressure, and loss of communication signals that can cause remote control execution failures during isolation and restoration processes. This study proposes an adaptive FLISR framework integrating graph theory with Deep Reinforcement Learning to address this gap, implemented on a modified test system based on the Sentarum Feeder at PLN UID West Kalimantan. A BreadthFirst Search (BFS) algorithm is developed to localize fault boundaries while automatically bypassing unhealthy devices, ensuring all isolation commands are physically executable by field actuators. The post-isolation service restoration is formulated as a Markov Decision Process with a tri-component state representation encoding topology, device health, and load restoration progress, optimized using a Double Deep Q-Network (DDQN) agent. A four-layer defense mechanism comprising BFS isolation, Service Restoration Pool filtering, Invalid-Action Masking, and health penalty in the reward function systematically excludes unhealthy assets from all FLISR decisions. Comparative analysis between the conventional approach and the proposed health-aware method demonstrates that the conventional BFS erroneously includes unhealthy devices in the isolation set leading to remote control failure, whereas the proposed method successfully bypasses such devices and identifies healthy alternatives. The simulation results confirm that the proposed framework determines optimal switch configuration while maintaining voltage profiles within 0.95-1.05 per unit and conductor loading below 80% of thermal capacity.