Scholarlink

<p>Water-intake pumping stations (WIPSs) are critical to waterworks energy consumption and operational costs. This paper addresses the optimal scheduling of multiple WIPS pumps, aiming to minimize total energy consumption while satisfying reservoir level, main pipe pressure difference, and pump switching frequency constraints. The problem is complicated by implicit system dynamics, temporally coupled constraints, and future water demand uncertainties. To tackle this, we reformulate it as a cooperative multi-agent Markov decision process (CM-MDP) with heterogeneous agents, and propose an energy-efficient scheduling algorithm by optimizing the integration of a surrogate model-assisted multi-agent attention-based deep reinforcement learning (SM-MAADRL) framework. Real-world simulations show the algorithm is highly robust to surrogate model uncertainties, reducing energy consumption by 2.04% compared to optimized Greedy strategies and 12.5% compared to standard MADDPG.</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

Multi-agent deep reinforcement learning for coordination scheduling of water-intake pumping stations

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Dongsheng Wang,
  • Ao Li,
  • Yujie Wei,
  • Jiahao Liu,
  • Liang Yu

摘要

Water-intake pumping stations (WIPSs) are critical to waterworks energy consumption and operational costs. This paper addresses the optimal scheduling of multiple WIPS pumps, aiming to minimize total energy consumption while satisfying reservoir level, main pipe pressure difference, and pump switching frequency constraints. The problem is complicated by implicit system dynamics, temporally coupled constraints, and future water demand uncertainties. To tackle this, we reformulate it as a cooperative multi-agent Markov decision process (CM-MDP) with heterogeneous agents, and propose an energy-efficient scheduling algorithm by optimizing the integration of a surrogate model-assisted multi-agent attention-based deep reinforcement learning (SM-MAADRL) framework. Real-world simulations show the algorithm is highly robust to surrogate model uncertainties, reducing energy consumption by 2.04% compared to optimized Greedy strategies and 12.5% compared to standard MADDPG.