Strategies are subject to certain limitations due to human origin. With global warming and other related phenomena, many unforeseen events are likely to occur. If operators do not have pre-established emergency plans for these scenarios, they will become overwhelmed and confused when dealing with malfunctions. The purpose of our study is to rapidly enhance strategy formulation to effectively respond to various unforeseen incidents. We introduce Agent NPP, a simulacrum of an NPP that include two types agents powered by LLMs: task executors and task validators. To integrate prior knowledge from probabilty safety analysis (PSA) to formulate strategies for NPPs, three tasks are designed: NPP initiating event sub-event analysis, event tree header event analysis, and strategy provision. They are closely related and collectively form an essential part of the NPP response system. Our central goal is to enable a strategy proposer agent to learn how to make emergency decisions within the simulacrum. Therefore, we propose a task-driven method called NPPAgent-Zero. The simulacrum can simulate the onset and progression of accidents based on knowledge bases of LLMs, allowing strategy proposer agents to accumulate experience from both successful and unsuccessful cases. This paper provides support for the rapid formulation of strategies.

LLM,decision support,nuclear power plant,agent,event tree anaysis