Scholarlink

<p>Daily pan evaporation is essential for water management in arid regions, but measurements are scarce and conventional methods require extensive inputs. This study presents a deployment-oriented framework for daily pan evaporation prediction using a compact long short-term memory (LSTM) model optimized by nature-inspired algorithms: Ant Colony Optimization, Grey Wolf Optimizer, and Whale Optimization Algorithm (WOA). Computational complexity is treated as a primary design criterion through explicit parameter counts and accuracy–capacity analysis. Application to three hyper-arid stations in Kuwait (Wafra, Kuwait International Airport, and Abdaly) shows close agreement with observations and limited systematic error across the full range. Across test-stage models, RMSE ranges from 1.1917 to 1.7491&#xa0;mm/day and <i>R</i><sup>2</sup> from 0.8544 to 0.9038. Performance is station-dependent: ACO performs best at Wafra and KIA, while Abdaly is better captured by persistence and Linear ARX. Results show that larger LSTM models do not guarantee better accuracy, and compact models should be preferred when they outperform simple baselines.</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

Complexity-Guided Modeling for Daily Pan Evaporation Prediction Using Compact LSTM Models and Nature-Inspired Optimization

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Abdullah A. Alsumaiei

摘要

Daily pan evaporation is essential for water management in arid regions, but measurements are scarce and conventional methods require extensive inputs. This study presents a deployment-oriented framework for daily pan evaporation prediction using a compact long short-term memory (LSTM) model optimized by nature-inspired algorithms: Ant Colony Optimization, Grey Wolf Optimizer, and Whale Optimization Algorithm (WOA). Computational complexity is treated as a primary design criterion through explicit parameter counts and accuracy–capacity analysis. Application to three hyper-arid stations in Kuwait (Wafra, Kuwait International Airport, and Abdaly) shows close agreement with observations and limited systematic error across the full range. Across test-stage models, RMSE ranges from 1.1917 to 1.7491 mm/day and R2 from 0.8544 to 0.9038. Performance is station-dependent: ACO performs best at Wafra and KIA, while Abdaly is better captured by persistence and Linear ARX. Results show that larger LSTM models do not guarantee better accuracy, and compact models should be preferred when they outperform simple baselines.