Model predictive control of DC/DC boost converter with reinforcement learning.

Area of Science:

• Power electronics and control systems engineering.
• Application of artificial intelligence in energy systems.

Background:

• Traditional Proportional-Integral (PI) controllers struggle with the inherent non-linearity of DC-DC converters, leading to transient instability and voltage fluctuations.
• The non-linear behavior of converters necessitates advanced control strategies that can adapt to varying operating conditions and loads.

Purpose of the Study:

• To investigate the efficacy of reinforcement learning (RL)-based non-linear controllers for DC-DC boost converters.
• To demonstrate the potential of RL controllers to improve control accuracy and energy efficiency compared to conventional methods.
• To develop a self-calibrating controller capable of optimizing its performance in real-time.

Main Methods:

• Utilizing a reinforcement learning model with a non-linear policy to minimize iteration and oscillation.
• Employing a support vector machine calibrated by RL to dynamically adjust the duty cycle limit under variable loads.
• Conducting simulation and experimental analyses on a DC-DC boost converter within a microgrid system.

Main Results:

• Reinforcement learning-based controllers show improved control and efficiency over standard controllers.
• The proposed RL controller effectively manages the dynamic changes in duty cycle limits required by variable loads.
• Demonstrated reduction in transient instability and voltage fluctuations compared to PI controllers.

Conclusions:

• Reinforcement learning offers a robust solution for controlling non-linear systems like DC-DC boost converters.
• Self-calibrating RL controllers provide adaptive and efficient power management, particularly in dynamic microgrid environments.
• The research validates the superiority of RL-based non-linear control for enhancing power electronic system performance.