Scholarlink

<p>Groundwater resources in arid and semi-arid regions such as Central Tunisia are under severe stress due to climate variability and escalating anthropogenic demands. Effective management of these resources requires advanced predictive tools capable of simulating aquifer responses to changing conditions. This study develops and evaluates a hybrid groundwater modeling framework for the Sbeitla Miocene Sandstone Aquifer by integrating the numerical model Visual MODFLOW with a data-driven Extreme Learning Machine (ELM) algorithm. The MODFLOW model was constructed using a single-layer, finite-difference grid (3000 cells) and calibrated under steady-state (2005) and transient (2005–2024) conditions, demonstrating strong performance (R² = 0.96 and 0.95, respectively). The ELM was trained on historical piezometric data to serve as a rapid surrogate model. Results indicate aquifer depletion, with water levels declining at average rates of 0.5&#xa0;m/year upstream and 0.5&#xa0;m/year in the central plain, culminating in total drawdowns of 15&#xa0;m over two decades. Pumping rates increased dramatically from 7.84&#xa0;Mm³/year to 16.23&#xa0;Mm³/year, far exceeding natural recharge 2.03&#xa0;Mm³/year, confirming excess groundwater abstraction. The ELM model outperformed MODFLOW in predictive accuracy, achieving an average R² of 0.97 and RMSE of 2.24&#xa0;m compared to MODFLOW’s R² of 0.95 and RMSE of 3.11&#xa0;m. Critically, residual analysis and Durbin-Watson (DW) statistics revealed that ELM residuals remained near-zero (mean − 0.03 to − 0.09&#xa0;m) with DW values consistently close to 2.0 (1.94–1.96), indicating no significant autocorrelation and statistically independent errors. In contrast, MODFLOW residuals showed progressive degradation (DW falling from 1.82 to 1.41), demonstrating systematic over-prediction and positive autocorrelation over time. Predictive simulations under a moderate climate change scenario (RCP 4.5) Predictive simulations under the Representative Concentration Pathway 4.5 (RCP 4.5), a moderate greenhouse gas stabilization scenario, indicate continued depletion through 2050, with an average additional drawdown of 11.7&#xa0;m and localized declines exceeding 17&#xa0;m, highlighting the urgent need for intervention. This hybrid approach combines the physical interpretability of MODFLOW with the computational efficiency and pattern-recognition strength of ELM, offering a robust tool for scenario analysis and sustainable resource management. This action plan would reduce the average drawdown to only 6&#xa0;m by 2050, thus preserving the resource for future generations while maintaining viable agricultural activity. The study underscores the critical need for implementing science-based management strategies, including abstraction limits and artificial recharge, to mitigate further depletion of this vital aquifer system. The hybrid ModFlow-ELM model is critically important because it uniquely combines MODFLOW’s physical interpretability and mass-balance rigor with ELM’s superior pattern-recognition speed and accuracy, offering a computationally efficient, robust tool for scenario analysis and sustainable groundwater management in arid and semi-arid regions.</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

Extreme Learning Machine (ELM) and MODFLOW, a hybrid model for groundwater flow modeling in the Sbeitla Miocene Sandstone Aquifer (Central Tunisia)

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Lamia Rachdi,
  • Hfidhi Awatef,
  • Maha Beji,
  • Amira Agoubi,
  • Mounir Atoui,
  • Bilel Abdelkrim,
  • Belgacem Agoubi

摘要

Groundwater resources in arid and semi-arid regions such as Central Tunisia are under severe stress due to climate variability and escalating anthropogenic demands. Effective management of these resources requires advanced predictive tools capable of simulating aquifer responses to changing conditions. This study develops and evaluates a hybrid groundwater modeling framework for the Sbeitla Miocene Sandstone Aquifer by integrating the numerical model Visual MODFLOW with a data-driven Extreme Learning Machine (ELM) algorithm. The MODFLOW model was constructed using a single-layer, finite-difference grid (3000 cells) and calibrated under steady-state (2005) and transient (2005–2024) conditions, demonstrating strong performance (R² = 0.96 and 0.95, respectively). The ELM was trained on historical piezometric data to serve as a rapid surrogate model. Results indicate aquifer depletion, with water levels declining at average rates of 0.5 m/year upstream and 0.5 m/year in the central plain, culminating in total drawdowns of 15 m over two decades. Pumping rates increased dramatically from 7.84 Mm³/year to 16.23 Mm³/year, far exceeding natural recharge 2.03 Mm³/year, confirming excess groundwater abstraction. The ELM model outperformed MODFLOW in predictive accuracy, achieving an average R² of 0.97 and RMSE of 2.24 m compared to MODFLOW’s R² of 0.95 and RMSE of 3.11 m. Critically, residual analysis and Durbin-Watson (DW) statistics revealed that ELM residuals remained near-zero (mean − 0.03 to − 0.09 m) with DW values consistently close to 2.0 (1.94–1.96), indicating no significant autocorrelation and statistically independent errors. In contrast, MODFLOW residuals showed progressive degradation (DW falling from 1.82 to 1.41), demonstrating systematic over-prediction and positive autocorrelation over time. Predictive simulations under a moderate climate change scenario (RCP 4.5) Predictive simulations under the Representative Concentration Pathway 4.5 (RCP 4.5), a moderate greenhouse gas stabilization scenario, indicate continued depletion through 2050, with an average additional drawdown of 11.7 m and localized declines exceeding 17 m, highlighting the urgent need for intervention. This hybrid approach combines the physical interpretability of MODFLOW with the computational efficiency and pattern-recognition strength of ELM, offering a robust tool for scenario analysis and sustainable resource management. This action plan would reduce the average drawdown to only 6 m by 2050, thus preserving the resource for future generations while maintaining viable agricultural activity. The study underscores the critical need for implementing science-based management strategies, including abstraction limits and artificial recharge, to mitigate further depletion of this vital aquifer system. The hybrid ModFlow-ELM model is critically important because it uniquely combines MODFLOW’s physical interpretability and mass-balance rigor with ELM’s superior pattern-recognition speed and accuracy, offering a computationally efficient, robust tool for scenario analysis and sustainable groundwater management in arid and semi-arid regions.