Loading...

Intelligent Control of DC-DC Power Electronic Converters Based on Deep Reinforcement Learning

Tavana, Ahmad | 2025

1 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58906 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Askarian, Iman
  7. Abstract:
  8. Decentralized control of DC microgrids plays a vital role in improving stability and transient performance in data centers, telecommunication systems, and renewable energy–based applications, all of which are considered critical infrastructures. In recent years, extensive research efforts have been conducted to enhance control strategies in this field, indicating that intelligent approaches can overcome the limitations of classical control methods. The main challenges in classical control include bus voltage restoration/regulation and proportional load current sharing among sources. These issues can lead to poor control performance under fast transient conditions, particularly in scenarios with variable loads. In this thesis, a model-free deep reinforcement learning–based controller is first designed and evaluated for regulating the output voltage of a flyback converter. Simulation results demonstrate that, under varying operating conditions and rapid load changes, the proposed controller provides faster transient response and improved voltage stability compared to a PI controller, while significantly reducing voltage overshoot. Subsequently, the proposed structure is extended to a multi-agent framework for droop control in a DC microgrid. This approach improves current sharing among sources and maintains DC bus voltage stability against disturbances without requiring any communication links. The advantages of this method include elimination of the need for an accurate system model, enhanced system security, and robustness against communication delays and unforeseen disturbances. One of the key features of the proposed control strategy is its scalability, such that adding or removing sources does not require reconfiguration or retraining of the agents. This property preserves effective control performance under structural changes in the microgrid. The results indicate that the proposed method provides a scalable solution with faster transient response and is well suited for decentralized control of DC microgrids
  9. Keywords:
  10. Droop Control ; Intelligent Control ; Deep Reinforcement Learning ; Multi-Agent Reinforcement Leaning ; Deep Q-Network (DQN)Algorithm ; Direct Current to Direct Current (DC to DC)Converters

 Digital Object List

 Bookmark

No TOC