Reinforcement Learning Based Coordinated Protection Using Conventional Relay Models

Abstract

This project investigates reinforcement learning for coordinated power system protection in a simulated 90 kV double-line transmission network. Conventional overcurrent, distance, and differential relay models are first implemented using synchronized voltage and current measurements. Instead of replacing these interpretable protection principles, a Deep Q-Network coordinator uses relay-derived features to select between waiting or relay-supported trip actions. The resulting framework combines physically meaningful relay evidence with learned coordination, aiming to reduce operating delay while maintaining selective and dependable fault isolation.