Scholarlink

<p>Groundwater level prediction in heterogeneous aquifer systems remains challenging because of hydroclimatic variability, high-dimensional inputs, and structural uncertainty. This study presents an integrated, uncertainty-aware framework that combines statistically grounded feature selection, irreducible noise estimation, and Bayesian adaptive model optimization for data-driven groundwater level forecasting. Two neural network architectures, namely a baseline multilayer perceptron (MLP) and an optimized MLP enhanced with the modified Chernobyl disaster optimizer (MLP–MCDO), were evaluated across 66 monitoring stations within Florida’s spatially heterogeneous aquifer system. Three complementary feature selection approaches, namely stability selection, entropy theory, and the Gamma test, were applied to identify robust predictors under differing statistical assumptions. A key innovation of this study is the application of Thompson sampling as a Bayesian multi-armed bandit framework for adaptively selecting optimal model–feature configurations under hydroclimatic uncertainty. By treating predictive configurations as stochastic alternatives and defining rewards through combined performance metrics, the framework enables probabilistic and station-specific model selection rather than static benchmarking. The Gamma test further provides an estimate of irreducible noise variance. It establishes a theoretical lower bound for prediction error and distinguishing meaningful system learning from overfitting. Analysis of groundwater variation rates and lagged hydroclimatic interactions revealed distinct regional control regimes. These include evapotranspiration-dominated dynamics in subtropical zones and rainfall-driven responses in temperate regions. The proposed framework offers a scalable and transferable approach for groundwater forecasting in complex environmental systems.</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

Enhancing Florida’s Groundwater Forecasting: A Hybrid Machine Learning Framework with Probabilistic Optimization and Smart Feature Selection

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Seyed Mostafa Biazar,
  • Golmar Golmohammadi,
  • Bhavan Voram,
  • Koroush Mohammadi

摘要

Groundwater level prediction in heterogeneous aquifer systems remains challenging because of hydroclimatic variability, high-dimensional inputs, and structural uncertainty. This study presents an integrated, uncertainty-aware framework that combines statistically grounded feature selection, irreducible noise estimation, and Bayesian adaptive model optimization for data-driven groundwater level forecasting. Two neural network architectures, namely a baseline multilayer perceptron (MLP) and an optimized MLP enhanced with the modified Chernobyl disaster optimizer (MLP–MCDO), were evaluated across 66 monitoring stations within Florida’s spatially heterogeneous aquifer system. Three complementary feature selection approaches, namely stability selection, entropy theory, and the Gamma test, were applied to identify robust predictors under differing statistical assumptions. A key innovation of this study is the application of Thompson sampling as a Bayesian multi-armed bandit framework for adaptively selecting optimal model–feature configurations under hydroclimatic uncertainty. By treating predictive configurations as stochastic alternatives and defining rewards through combined performance metrics, the framework enables probabilistic and station-specific model selection rather than static benchmarking. The Gamma test further provides an estimate of irreducible noise variance. It establishes a theoretical lower bound for prediction error and distinguishing meaningful system learning from overfitting. Analysis of groundwater variation rates and lagged hydroclimatic interactions revealed distinct regional control regimes. These include evapotranspiration-dominated dynamics in subtropical zones and rainfall-driven responses in temperate regions. The proposed framework offers a scalable and transferable approach for groundwater forecasting in complex environmental systems.