Scholarlink

<p>Evaporation is a crucial part of the water cycle and is significantly influenced by weather conditions, particularly temperature. In this study, we evaluated the performance of seven machine learning models in predicting monthly pan evaporation (EP) (mm) across five reservoir dams in northeastern Algeria, which vary significantly in design, hydro-climatic conditions, and operational purposes. The models included Artificial Neural Networks (ANN), Classification and Regression Trees (CART), K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Machines (SVM), as well as RF and SVM optimized using Particle Swarm Optimization (PSO), resulting in the hybrid RF-PSO and SVM-PSO models. The methods were trained using air temperature (°C), relative humidity (%), and wind speed (m/s) characteristics of the study dam reservoirs, and were evaluated using five performance metrics (RMSE, MAE, NSE, and R<sup>2</sup>) and visually compared with Taylor diagrams. All models showed good performance, with RF and RF-PSO consistently providing the highest accuracy. The results showed that RF-PSO and RF consistently outperformed other models across all five reservoirs, with testing RMSE ranging from 0.038 mm to 0.112 mm, R<sup>2</sup> values between 0.737 and 0.983, and NSE typically exceeding 0.9, demonstrating strong robustness and generalization. SVM also performed reliably, reaching R<sup>2</sup> values up to 0.985 in some sites. In contrast, the hybrid SVM-PSO model, despite achieving excellent training accuracy, showed unstable testing performance in several reservoirs, indicating sensitivity to PSO-based parameter tuning. ANN and CART achieved moderate accuracy (testing R<sup>2</sup> = 0.648–0.957; NSE = 0.615–0.946), while KNN exhibited pronounced overfitting, achieving perfect performance during training (RMSE = 0; R<sup>2</sup> = 1.0; NSE = 1.0) but experiencing a considerable decline in performance during the testing phase. By applying machine learning models across climatically and structurally diverse reservoirs, our study demonstrated the superiority and adaptability of two RF variants, the single RF and the hybrid RF-PSO models, for EP modeling in heterogeneous environments. The results highlight the need for site-specific calibration and provide a methodological foundation to support water resource management, dam operation, and evaporation estimation under variable and data-scarce climatic conditions.</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

Comparing the Performance of Single and Hybrid Machine Learning Models for Estimating Pan Evaporation Across Climatically and Structurally Diverse Reservoirs

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Boukemache Khaoula,
  • Marouf Nadir,
  • Batout Soumia,
  • Binh Thai Pham

摘要

Evaporation is a crucial part of the water cycle and is significantly influenced by weather conditions, particularly temperature. In this study, we evaluated the performance of seven machine learning models in predicting monthly pan evaporation (EP) (mm) across five reservoir dams in northeastern Algeria, which vary significantly in design, hydro-climatic conditions, and operational purposes. The models included Artificial Neural Networks (ANN), Classification and Regression Trees (CART), K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Machines (SVM), as well as RF and SVM optimized using Particle Swarm Optimization (PSO), resulting in the hybrid RF-PSO and SVM-PSO models. The methods were trained using air temperature (°C), relative humidity (%), and wind speed (m/s) characteristics of the study dam reservoirs, and were evaluated using five performance metrics (RMSE, MAE, NSE, and R2) and visually compared with Taylor diagrams. All models showed good performance, with RF and RF-PSO consistently providing the highest accuracy. The results showed that RF-PSO and RF consistently outperformed other models across all five reservoirs, with testing RMSE ranging from 0.038 mm to 0.112 mm, R2 values between 0.737 and 0.983, and NSE typically exceeding 0.9, demonstrating strong robustness and generalization. SVM also performed reliably, reaching R2 values up to 0.985 in some sites. In contrast, the hybrid SVM-PSO model, despite achieving excellent training accuracy, showed unstable testing performance in several reservoirs, indicating sensitivity to PSO-based parameter tuning. ANN and CART achieved moderate accuracy (testing R2 = 0.648–0.957; NSE = 0.615–0.946), while KNN exhibited pronounced overfitting, achieving perfect performance during training (RMSE = 0; R2 = 1.0; NSE = 1.0) but experiencing a considerable decline in performance during the testing phase. By applying machine learning models across climatically and structurally diverse reservoirs, our study demonstrated the superiority and adaptability of two RF variants, the single RF and the hybrid RF-PSO models, for EP modeling in heterogeneous environments. The results highlight the need for site-specific calibration and provide a methodological foundation to support water resource management, dam operation, and evaporation estimation under variable and data-scarce climatic conditions.