Development of Autonomous Operation Agent for Normal and Emergency Situations in Nuclear Power Plants

Nuclear power plants (NPPs) are highly automated systems. Nevertheless, the operator’s manual actions are still required when normal (start-up/shutdown) operation and emergency operation. During these operations, operators execute necessary actions (e.g., situation awareness, confirmation of automatic actuation, and manipulation) following the operating procedures. This study suggests a Deep Reinforcement Learning (DRL)-based autonomous agent. The agent can manage the power increase operation from 2% to 100% and reduce the pressure and temperature until the shutdown cooling entry condition after reactor trip caused by loss of coolant accident in NPPs. The DRL-based controller suggested in this study combines a rule-based system and DRL algorithm that involves a Soft Actor-Critic algorithm and deep neural network. The test results using a compact nuclear simulator indicates that the agent can manipulate components to comply with identified constraints for start-up operation and emergency operation.