Scholarlink

<p>This study evaluates the predictive capabilities of multiple machine learning (ML) models and an ensemble stacking model (XGBoost) across different input-feature scenarios. Our results demonstrate that solely relying on water level measurements from neighboring wells provides limited predictive value, whereas incorporating inter-well distances significantly improves accuracy. For instance, the Gradient Boosting Regressor (GBR) testing RMSE improves from 23.5&#xa0;m (with 3 input features) to 15.1&#xa0;m (with 10 input features). Among individual models, Random Forest (RF) and GBR achieve the best generalization, while the stacking ensemble model (extreme gradient boosting: XGBoost) consistently outperforms all base models. To further enhance model efficiency, we explore five different dimensionality reduction techniques. Principal Component Analysis (PCA) emerges as the most suitable method, effectively reducing dimensionality while preserving variance. After applying PCA to a 10-feature dataset (levels and distances from five neighboring wells), model performance remains strong, with XGBoost maintaining an R² of 0.91 and a slight RMSE increase to 8.1&#xa0;m. These findings underscore the importance of incorporating spatial relationships in groundwater modeling and highlight the benefits of ensemble learning for robust predictions. By integrating ML with dimensionality reduction, this study provides a scalable, data-driven framework for groundwater level modeling, supporting improved water management strategies in complex hydrogeological settings.</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

Evaluating the efficacy of PCA and t-SNE in optimizing input features for groundwater level simulation using machine learning models

  • 期刊论文

  • 发表日期：

    2025年6月2日

查看全文
  • Nan Wang,
  • Qiang Zhou,
  • Jinhua Gao,
  • Zi’xi Wang

摘要

This study evaluates the predictive capabilities of multiple machine learning (ML) models and an ensemble stacking model (XGBoost) across different input-feature scenarios. Our results demonstrate that solely relying on water level measurements from neighboring wells provides limited predictive value, whereas incorporating inter-well distances significantly improves accuracy. For instance, the Gradient Boosting Regressor (GBR) testing RMSE improves from 23.5 m (with 3 input features) to 15.1 m (with 10 input features). Among individual models, Random Forest (RF) and GBR achieve the best generalization, while the stacking ensemble model (extreme gradient boosting: XGBoost) consistently outperforms all base models. To further enhance model efficiency, we explore five different dimensionality reduction techniques. Principal Component Analysis (PCA) emerges as the most suitable method, effectively reducing dimensionality while preserving variance. After applying PCA to a 10-feature dataset (levels and distances from five neighboring wells), model performance remains strong, with XGBoost maintaining an R² of 0.91 and a slight RMSE increase to 8.1 m. These findings underscore the importance of incorporating spatial relationships in groundwater modeling and highlight the benefits of ensemble learning for robust predictions. By integrating ML with dimensionality reduction, this study provides a scalable, data-driven framework for groundwater level modeling, supporting improved water management strategies in complex hydrogeological settings.