Scholarlink

<p>This study conducts a large-scale analysis across the state of Florida in the United States, systematically evaluating the efficacy of four advanced deep learning models (Long Short-Term Memory (LSTM), Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS), Neural Hierarchical Interpolation for Time Series (N-HiTS), and Transformer) using extensive datasets from 45 surface water and groundwater observation stations, spanning 23 years (2001–2023). After rigorous data preprocessing, these models were trained and tested to predict daily water dynamics. To enhance interpretability and trust in model predictions, the SHapley Additive Explanations (SHAP) was applied to quantify the influence of input variables on model outputs. SHAP quantified feature contributions and enhanced interpretability. N-HiTS outperformed others in 70% of groundwater stations (RMSE = 0.183&#xa0;m; 14% lower than LSTM, 26% lower than Transformer) and 30% of surface water stations. N-BEATS excelled in 50% of surface water stations (NRMSE = 0.052; 11% improvement over LSTM, 27% over Transformer). SHAP revealed 7-day lagged target variables contributed 45–68% of predictions, confirming strong memory effects in Florida’s karst systems, with soil temperature, solar radiation, and rainfall as key secondary drivers, delivering actionable hydrological insights.</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

Leveraging advanced deep learning models for large-scale hydrological forecasting in Florida

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Saba Shaghaghi,
  • Golmar Golmohammadi,
  • Saman Javadi,
  • Seyed Mostafa Biazar,
  • Rohith Reddy Nedhunuri,
  • Kourosh Mohammadi,
  • Haimanote Bayabil,
  • Davie M. Kadyampakeni,
  • Mary G. Lusk

摘要

This study conducts a large-scale analysis across the state of Florida in the United States, systematically evaluating the efficacy of four advanced deep learning models (Long Short-Term Memory (LSTM), Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS), Neural Hierarchical Interpolation for Time Series (N-HiTS), and Transformer) using extensive datasets from 45 surface water and groundwater observation stations, spanning 23 years (2001–2023). After rigorous data preprocessing, these models were trained and tested to predict daily water dynamics. To enhance interpretability and trust in model predictions, the SHapley Additive Explanations (SHAP) was applied to quantify the influence of input variables on model outputs. SHAP quantified feature contributions and enhanced interpretability. N-HiTS outperformed others in 70% of groundwater stations (RMSE = 0.183 m; 14% lower than LSTM, 26% lower than Transformer) and 30% of surface water stations. N-BEATS excelled in 50% of surface water stations (NRMSE = 0.052; 11% improvement over LSTM, 27% over Transformer). SHAP revealed 7-day lagged target variables contributed 45–68% of predictions, confirming strong memory effects in Florida’s karst systems, with soil temperature, solar radiation, and rainfall as key secondary drivers, delivering actionable hydrological insights.