Scholarlink

<p>Irrigation networks in arid and semi-arid regions face water deficits and distribution losses, making reliable operation a challenge. This study investigates the Dez distributary E1R1 (Iran) with three pool-controlled reaches (Ch1–Ch3) and six turnouts (TO1–TO6) using operator records and Irrigation Conveyance System Simulation (ICSS) scenarios. Objectives are to infer an operator-consistent reward using Maximum Entropy Inverse Reinforcement Learning (MEIRL), derive gate-control policies from the recovered reward, and evaluate performance using indicators. The framework couples MEIRL with an ICSS Saint–Venant simulator, learning from expert state–action demonstrations and testing the policy across scenarios. Unlike canal-control reinforcement learning that prescribes or tunes a reward (e.g., fuzzy SARSA), this approach infers a canal-specific reward from expert behavior, yielding an auditable preference model for policy derivation and validation. Water-level regulation is assessed via maximum and integral absolute error (MAE, IAE), while delivery service is assessed via Molden–Gates indicators: efficiency (MPF), equity (MPE), and dependability (MPD). After learning, MAE falls below 5% (vs. &gt; 20% during learning). Post-learning, Ch1 remains stable (MAE 1.9–4.6%), Ch2 shows residual volatility (peak MAE 14.3%), and Ch3 is precise (MAE ≤ 1.3% in most scenarios; one at 12.9%); mean IAE values are 1.65% (Ch1), 4.34% (Ch2), and 1.64% (Ch3). MPF remains within 0.85–1.0 (minimum 0.951 at TO3), with improved dispersion (MPE 0.001–0.03; MPD 0.001–0.023). Results provide an auditable pathway for encoding operational preferences into deployable canal policies, motivating persistence-aware rewards, scaling to larger networks, and indicator-based auditing for data-scarce commands.</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

Expert-driven canal control using inverse reinforcement learning for minimizing water level and delivery errors in irrigation networks

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Kazem Shahverdi,
  • Hesam Ghodousi,
  • Atefeh Norouzi,
  • Ahmed M. Saqr

摘要

Irrigation networks in arid and semi-arid regions face water deficits and distribution losses, making reliable operation a challenge. This study investigates the Dez distributary E1R1 (Iran) with three pool-controlled reaches (Ch1–Ch3) and six turnouts (TO1–TO6) using operator records and Irrigation Conveyance System Simulation (ICSS) scenarios. Objectives are to infer an operator-consistent reward using Maximum Entropy Inverse Reinforcement Learning (MEIRL), derive gate-control policies from the recovered reward, and evaluate performance using indicators. The framework couples MEIRL with an ICSS Saint–Venant simulator, learning from expert state–action demonstrations and testing the policy across scenarios. Unlike canal-control reinforcement learning that prescribes or tunes a reward (e.g., fuzzy SARSA), this approach infers a canal-specific reward from expert behavior, yielding an auditable preference model for policy derivation and validation. Water-level regulation is assessed via maximum and integral absolute error (MAE, IAE), while delivery service is assessed via Molden–Gates indicators: efficiency (MPF), equity (MPE), and dependability (MPD). After learning, MAE falls below 5% (vs. > 20% during learning). Post-learning, Ch1 remains stable (MAE 1.9–4.6%), Ch2 shows residual volatility (peak MAE 14.3%), and Ch3 is precise (MAE ≤ 1.3% in most scenarios; one at 12.9%); mean IAE values are 1.65% (Ch1), 4.34% (Ch2), and 1.64% (Ch3). MPF remains within 0.85–1.0 (minimum 0.951 at TO3), with improved dispersion (MPE 0.001–0.03; MPD 0.001–0.023). Results provide an auditable pathway for encoding operational preferences into deployable canal policies, motivating persistence-aware rewards, scaling to larger networks, and indicator-based auditing for data-scarce commands.