2D Iterative Learning Control with Deep Reinforcement Learning
Compensation for the Non-repetitive Batch Processes
Abstract
Iterative learning control (ILC) is an advantage control strategy widely
used in batch systems. Nevertheless, designing an effective iterative
learning control scheme is still a critical problem for complex batch
processes with non-repetitive nature and model mismatch. In this paper,
we propose a two-dimensional iterative learning control-reinforcement
learning (2D ILC-RL) control scheme composed of a two-dimensional ILC
controller and a two-dimensional DRL compensator. Based on the 2D system
theory, the 2D ILC controller is proposed to ensure the primary control
performance and its stability and convergence are verified. Meanwhile,
the DRL compensator counteracts the negative impact of the model
mismatch and the non-repetitive nature. In addition, we proposed a
real-time implementation scheme to guarantee the safety of the practical
batch systems compared to the conventional online training method.
Finally, the simulation results in two chemical batch processes
demonstrate the proposed control method’s effectiveness, significant
control performance, and strong robustness.