Research And Simulation Of Reinforcement Learning Control Strategy For Nuclear Reactor Primary SGTR

Due to the increase in energy demand and environmental pollution caused by the consumption of fossil fuels because of economic development driven by globalization,countries are moving towards energy conservation,emission reduction,and the development of clean and efficient new energy.Nuclear energy is a low-carbon and stable energy source that can reduce air pollution,optimize the structure of the power industry,and achieve sustainable development.To promote autonomous monitoring technology based on artificial intelligence in nuclear reactors,China’s "13th Five Year Plan" nuclear power development project aims to create a prototype to achieve highly autonomous execution of typical operational tasks and accident handling processes.By using model recognition and reinforcement learning,offline systems can be identified within a certain range of expected accidents,and adaptive controllers for autonomous control can be designed.This research focuses on using reinforcement learning to complete complex control functions during typical shutdown responses,thereby improving the intelligence of the unit under accident conditions.The specific research content of this article is as follows:(1)The application of artificial intelligence in reactor control is analyzed,and existing fault handling methods and artificial intelligence methods are explained.This paper introduces the important impact of SGTR(steam generator heat transfer tube rupture)on the reactor and the operation process of the accident on a simulator.It introduces and analyzes the basic theory and strategy of time series modeling,and selects parameters based on the characteristics and characteristic analysis of the reactor accident process.Provide data support for subsequent research.(2)The correlation between primary circuit temperature parameters,pressurizer water level and pressure,and other parameters during SGTR accidents was analyzed.The LSTM neural network framework has been adjusted to apply to reinforcement learning.Based on this,a primary circuit temperature rate model and a pressurizer pressure and level model suitable for reinforcement learning have been established,laying the foundation for the application of deep reinforcement learning in reactor accident control.(3)The basic theory of deep reinforcement learning is presented,and a process control system framework and learning framework based on deep reinforcement learning under fault conditions are proposed.Based on this,a simulation experiment is conducted using TD3 algorithm to control the temperature change rate of the reactor primary circuit,which proves the effectiveness of the algorithm.Aiming at the multi-input multi output problem of reactor pressurizer under fault state,it is shown that the PPO algorithm is more suitable for multi objective control.Simulation experiments show that the control system has good set point tracking performance.