Reinforcement learning based adaptive control system and method for nuclear fusion

Abstract

The application discloses a kind of nuclear fusion adaptive control system and method based on reinforcement learning, the control system includes physical layer and control layer, physical layer contains tokamak / stellarator device, Thomson scattering and so on diagnostic system and FPAD / PAM projectile injection system, responsible for collecting plasma density profile, temperature profile and other data and executing fuel supply;Control layer is reinforcement learning intelligent agent, contains state perception, strategy network and execution interface module, strategy network uses PPO / SAC algorithm, state perception is included line average density, density change rate and historical sequence data;The control method is trained offline by Torax simulator, is optimized by composite reward function, forms perception-decision-making-execution-feedback closed-loop control, supports real machine deployment and online fine tuning.The application solves the problem of traditional PID control lag, adaptive difference, the control precision is improved by about 40%, density fluctuation is reduced by about 50%, prolongs hardware maintenance cycle, adapts nuclear fusion long pulse operation demand.

Images