| description abstract | Frequent changes in operating conditions can result in the great loss of the service life of gas turbines that work at high speed, high pressure, and high temperature. To improve the control system of gas turbines is of great significance for extending the service life, boosting the dynamic performance, and reducing the maintenance cost. Due to good dynamic response characteristics, the present control methods are neither capable of tackling the nonlinearity of the system nor adaptive to the frequent variations of operating conditions. As a powerful learning paradigm, reinforcement learning (RL) can explore the operation environment and make decisions adaptively. A novel intelligent control framework for the gas turbine control system is constructed by coupling a RL agent with a dynamic simulation model and a damage estimation model. Compared with the proportion, integral, differential (PID) and fuzzy control, the proposed method achieves better performance in both extending life and depicting dynamic characteristics, and reduces the overall damage to as low as 0.01% in the loading process. Besides, the overshoot and the adjusting time of the novel approach are lower and shorter than those of PID and fuzzy control by more than 90% and about 14%, respectively, but it takes longer to accelerate. Finally, the effect of different types of RL agents and their hyperparameters are investigated. The results show that the deep deterministic policy gradient (DDPG) agent can achieve the best performance. Furthermore, in addition to extending the life and improving the dynamic performance, the controlling framework presented is recommended to construct an intelligent dynamic control system for achieving various purposes. | |
