Multivariable Coupled System Control Method Based on Deep Reinforcement Learning

Abstract

Due to the multi-loop coupling characteristics of multivariable systems, it is difficult for traditional control methods to achieve precise control effects. Therefore, this paper proposes a control method based on deep reinforcement learning to achieve stable and accurate control of multivariable coupling systems. Based on the proximal policy optimization algorithm (PPO), this method selects tanh as the activation function and normalizes the advantage function. At the same time, based on the characteristics of the multivariable coupling system, the reward function and controller are redesigned structures, achieving stable and precise control of the controlled system. In addition, this study used the amplitude of the control quantity output by the controller as an indicator to evaluate the controller’s performance. Finally, simulation verification was conducted in MATLAB/Simulink. The experimental results show that compared with decentralized control, decoupled control and traditional PPO control, the method proposed in this article achieves better control effects.