One policy to rule them all: Handling multiple emergent accidents in nuclear power plants with ensemble-based behavior cloning

Nuclear energy plays a vital role in the global shift toward sustainable power. Ensuring the safety and efficiency of nuclear power plant (NPP) operations is critically important, especially in the presence of multiple emergent accident scenarios. This highlights a pressing need for a controller capable of automatically and rapidly responding to multiple emergent accidents. To address this, we present a Markov Decision Process (MDP) model tailored to the nuclear power plant environment, focusing on six representative accident conditions. Building on this foundation, we introduce two novel algorithms – Ensemble-based Cross-Task Behavior Cloning (ECT-BC) and Ensemble-based Cross-Task Value-Weighted Behavior Cloning (ECT-VWBC) – which integrate Deep Reinforcement Learning (DRL) and Ensemble Learning to train an agent for handling multiple emergent accidents in NPPs. ECT-BC employs offline imitation learning to train policy models using expert data of numerous emergent accidents simultaneously. ECT-VWBC further introduces a value-based weighting scheme that prioritizes training samples with higher estimated action values, enhancing the influence of optimal actions. We propose ensemble-based methods with a voting mechanism to enhance policy diversity and generalization. Experimental results on the Nuclear Power Plant Accident Conditions MDP demonstrate the effectiveness of our algorithms, providing a foundation for more reliable autonomous decision-making systems in nuclear safety applications.