Scholarlink

<p>Adaptive reservoir management is essential for semi-arid regions facing heightened climate variability and intensifying competition among water use sectors. Traditional rule-based and optimization-based approaches frequently lack the responsiveness needed to address hydrological extremes and shifting operational priorities. To address this gap, we present a systematic comparison of classical and artificial intelligence-based reservoir operation models under diverse hydrological scenarios. Four strategies—Linear Programming (LP), Stochastic Optimization (SO), Reinforcement Learning (RL), and Hierarchical Reinforcement Learning (HRL)—were evaluated for the Oued Makhazine Dam in northern Morocco using a combination of historical records and synthetically generated inflow and demands sequences simulating dry, historical, and rainy regimes. Model performance was assessed using comprehensive criteria, including allocation efficiency, sectoral demand satisfaction, safety margin, adaptability, reliability index, and the frequency of reservoir volume extremes. Results indicate that the SO model demonstrated superior overall performance under stable conditions, achieving the highest efficiency (72.13 ± 37.70%), irrigation satisfaction (92.43 ± 10.59%), drinking water satisfaction (99.63 ± 0.51%), and safety margin (55.84 ± 37.23%). While SO maintained a high average buffer, it also frequently reached maximum capacity during rainy periods, increasing the risk of overflow. HRL excelled in adaptability (82.09 ± 12.51%) and consistently ensured drinking water supply (98.02 ± 2.80%). RL showed high adaptability (82.30 ± 11.89%), indicating strong responsiveness to fluctuating demands. However, this flexibility came at the cost of greater reservoir volume variability (105.38 ± 56.43) and inconsistent drinking water satisfaction (84.67 ± 26.22%), particularly under dry conditions. LP achieved high satisfaction rates for drinking (96.97 ± 4.29%) and irrigation (77.50 ± 29.98%) but was less adaptable (71.15 ± 0.64%), limiting its performance under variable regimes.These findings underscore the advantages of SO under predictable conditions and highlight the adaptive potential of RL-based approaches for dynamic water management. The study contributes to the evaluation of reservoir operation strategies under climate uncertainty and supports the design of adaptive methodologies for sustainable water resource management in semi-arid environments.</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

Comparative Analysis of Water Management Strategies at Oued Makhazine Dam: Assessing the Performance of Reinforcement Learning and Traditional Optimization Methods

  • 期刊论文

  • 发表日期：

    2025年9月25日

查看全文
  • Chakir Achahboun,
  • Mohamed Chikhaoui,
  • Mustapha Naimi,
  • Mostafa Bellafkih

摘要

Adaptive reservoir management is essential for semi-arid regions facing heightened climate variability and intensifying competition among water use sectors. Traditional rule-based and optimization-based approaches frequently lack the responsiveness needed to address hydrological extremes and shifting operational priorities. To address this gap, we present a systematic comparison of classical and artificial intelligence-based reservoir operation models under diverse hydrological scenarios. Four strategies—Linear Programming (LP), Stochastic Optimization (SO), Reinforcement Learning (RL), and Hierarchical Reinforcement Learning (HRL)—were evaluated for the Oued Makhazine Dam in northern Morocco using a combination of historical records and synthetically generated inflow and demands sequences simulating dry, historical, and rainy regimes. Model performance was assessed using comprehensive criteria, including allocation efficiency, sectoral demand satisfaction, safety margin, adaptability, reliability index, and the frequency of reservoir volume extremes. Results indicate that the SO model demonstrated superior overall performance under stable conditions, achieving the highest efficiency (72.13 ± 37.70%), irrigation satisfaction (92.43 ± 10.59%), drinking water satisfaction (99.63 ± 0.51%), and safety margin (55.84 ± 37.23%). While SO maintained a high average buffer, it also frequently reached maximum capacity during rainy periods, increasing the risk of overflow. HRL excelled in adaptability (82.09 ± 12.51%) and consistently ensured drinking water supply (98.02 ± 2.80%). RL showed high adaptability (82.30 ± 11.89%), indicating strong responsiveness to fluctuating demands. However, this flexibility came at the cost of greater reservoir volume variability (105.38 ± 56.43) and inconsistent drinking water satisfaction (84.67 ± 26.22%), particularly under dry conditions. LP achieved high satisfaction rates for drinking (96.97 ± 4.29%) and irrigation (77.50 ± 29.98%) but was less adaptable (71.15 ± 0.64%), limiting its performance under variable regimes.These findings underscore the advantages of SO under predictable conditions and highlight the adaptive potential of RL-based approaches for dynamic water management. The study contributes to the evaluation of reservoir operation strategies under climate uncertainty and supports the design of adaptive methodologies for sustainable water resource management in semi-arid environments.