Scholarlink

<p>This paper introduces a Deep Reinforcement Learning (DRL) approach for real-time energy management in drones, aimed at optimizing the trade-off between battery power and energy harvested from piezoelectric sensors. By modeling the problem as a Markov Decision Process (MDP), the proposed system enables drones to make context-aware decisions based on current environmental and energy states, favoring energy-efficient actions such as switching to harvested power when feasible. Experimental results show that the DRL-based model extends flight duration by up to 20%, improves energy efficiency by 18%, and achieves over 95% decision accuracy across diverse operating conditions. These performance gains significantly surpass both rule-based and heuristic strategies in enhancing drone autonomy. Despite these promising outcomes, real-world deployment remains limited by simulation-based assumptions, delayed reactions to rare events, and computational demands. Future research will focus on model refinement, real-world adaptation via transfer learning, and the development of more realistic simulation environments to strengthen the practical viability of DRL-driven energy management in autonomous drones.</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

Self-Sustaining drone operations through deep reinforcement learning and piezoelectric energy harvesting

  • 期刊论文

  • 发表日期：

    2025年8月18日

查看全文
  • Abderaouf Bahi,
  • Amel Ourici

摘要

This paper introduces a Deep Reinforcement Learning (DRL) approach for real-time energy management in drones, aimed at optimizing the trade-off between battery power and energy harvested from piezoelectric sensors. By modeling the problem as a Markov Decision Process (MDP), the proposed system enables drones to make context-aware decisions based on current environmental and energy states, favoring energy-efficient actions such as switching to harvested power when feasible. Experimental results show that the DRL-based model extends flight duration by up to 20%, improves energy efficiency by 18%, and achieves over 95% decision accuracy across diverse operating conditions. These performance gains significantly surpass both rule-based and heuristic strategies in enhancing drone autonomy. Despite these promising outcomes, real-world deployment remains limited by simulation-based assumptions, delayed reactions to rare events, and computational demands. Future research will focus on model refinement, real-world adaptation via transfer learning, and the development of more realistic simulation environments to strengthen the practical viability of DRL-driven energy management in autonomous drones.