In this work, a reinforcement learning-aided proportional-integral-derivative control algorithm is developed. This algorithm uses Q-learning as a supervisory layer to learn the controller tuning parameters. Since existing algorithms in this area can lead to intractability due the growing Q-table for a high-dimensional state-space system, a clustering technique is proposed to shrink the state space thus leading to a computationally-tractable problem. The algorithm is applied to the control of the main steam temperature in a supercritical pulverized coal power plant. For this highly nonlinear system, it is observed that the RL-aided coordinated control strategy results in a superior control performance in comparison to the traditional static controllers. The RL algorithm can be readily applied to any existing control system with minimum disruption in the control structure.
展开▼