Nonlinear Control of a Gas Turbine Engine with Reinforcement Learning

Inspired by the recent progress in intelligent learning, in this work, a model-free reinforcement learning (RL) control strategy is developed to improve the shaft speed response of a laboratory gas turbine engine . The primary objective of the RL-based control system is to efficiently track the reference signal of shaft speed while adhering to constraints with minimal control effort. A model-free, deterministic policy-based reinforcement learning algorithm, viz., the Actor-Critic (A2C) algorithm, is used for the control design. The learning is through direct interaction with the system without relying on the mathematical model dynamics of the system. The proposed A2C-RL control scheme’s performance is compared with that of a well-tuned proportional-integral-derivative (PID) controller. The resulting closed-loop responses show that the proposed A2C-RL controller outperforms the PID controller in achieving better transient performance with less control effort.