Scholarlink

<p>This paper proposes a reinforcement learning-based optimal control strategy for proton exchange membrane fuel cells (PEMFCs) with the objective of maximizing energy conversion efficiency and improving adaptability under varying operating conditions. The control problem is formulated as a Markov decision process and addressed using Q-learning and deep Q-network (DQN) algorithms. Appropriate state, action, and reward structures are designed to enable effective learning of the maximum power point (MPP) without requiring explicit knowledge of system dynamics. The proposed controllers are evaluated through extensive simulations under temperature variations, membrane water content changes, and dynamic load conditions. Simulation results show that DQN achieves approximately twice the learning speed of Q-learning while maintaining comparable MPP performance. Compared with traditional model-based and heuristic methods such as Perturb and Observe, the reinforcement learning-based approaches demonstrate superior adaptability to system nonlinearities and robustness under dynamic operating scenarios. Furthermore, DQN exhibits the fastest convergence time of 1.30 s while reaching an MPP of 6477.14 W, confirming its effectiveness for real-time PEMFC control applications. In comparison, the convergence time and MPP value of Q-learning and P&amp;O based on these scenarios was found to be 3.70 s–6477.14W, and 2.70 s–6467.33W, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Performance improvement of proton exchange membrane fuel cells via deep Q-network

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Duong Tuan Nguyen,
  • Nhat-Minh Le-Phan,
  • Nga Thi-Thuy Vu

摘要

This paper proposes a reinforcement learning-based optimal control strategy for proton exchange membrane fuel cells (PEMFCs) with the objective of maximizing energy conversion efficiency and improving adaptability under varying operating conditions. The control problem is formulated as a Markov decision process and addressed using Q-learning and deep Q-network (DQN) algorithms. Appropriate state, action, and reward structures are designed to enable effective learning of the maximum power point (MPP) without requiring explicit knowledge of system dynamics. The proposed controllers are evaluated through extensive simulations under temperature variations, membrane water content changes, and dynamic load conditions. Simulation results show that DQN achieves approximately twice the learning speed of Q-learning while maintaining comparable MPP performance. Compared with traditional model-based and heuristic methods such as Perturb and Observe, the reinforcement learning-based approaches demonstrate superior adaptability to system nonlinearities and robustness under dynamic operating scenarios. Furthermore, DQN exhibits the fastest convergence time of 1.30 s while reaching an MPP of 6477.14 W, confirming its effectiveness for real-time PEMFC control applications. In comparison, the convergence time and MPP value of Q-learning and P&O based on these scenarios was found to be 3.70 s–6477.14W, and 2.70 s–6467.33W, respectively.