Multivariable Coupled System Control Method Based on Deep Reinforcement Learning

Area of Science:

• Control Engineering
• Artificial Intelligence
• Systems Science

Background:

• Traditional control methods struggle with precise control of multivariable systems due to multi-loop coupling.
• Developing advanced control strategies is crucial for improving system performance and stability.

Purpose of the Study:

• To propose a novel deep reinforcement learning-based control method for multivariable coupling systems.
• To achieve stable and accurate control beyond the capabilities of existing techniques.

Main Methods:

• Utilized the proximal policy optimization (PPO) algorithm with tanh activation and normalized advantage function.
• Redesigned reward function and controller structures tailored to multivariable coupling system characteristics.
• Evaluated controller performance using the amplitude of the control quantity output.

Main Results:

• The proposed deep reinforcement learning method demonstrated superior control effects compared to decentralized control, decoupled control, and traditional PPO.
• Simulation verification in MATLAB/Simulink confirmed the enhanced stability and precision of the new control strategy.
• The redesigned reward function and controller structure effectively addressed the challenges of multivariable coupling.

Conclusions:

• The deep reinforcement learning approach offers a significant advancement for controlling complex multivariable systems.
• This method provides a robust and accurate solution for systems previously difficult to manage with conventional techniques.
• The study highlights the potential of tailored deep reinforcement learning for sophisticated engineering control applications.