Scholarlink

<p>Accurate and rapid forecasting of nearshore water levels is crucial for managing estuarine and coastal ecosystems, where tidal dynamics, river discharge, and human activities interact in complex ways. This study employs AutoGluon, an automated ensemble machine learning framework, to predict tidal water levels in the Pearl River Delta and compares its performance against a Long Short-Term Memory (LSTM) model under diverse tidal conditions. Our results demonstrate that the proposed framework achieves predictive accuracy comparable to the LSTM baseline, as evaluated by the Nash-Sutcliffe Efficiency (NSE), while reducing model training time by nearly sevenfold. The potential of transfer learning for cross-regional prediction is further assessed to address data scarcity in ungauged areas. The transferred model attains NSE values exceeding 0.85 at unseen stations without prior local training, improving prediction accuracy by an average of 15.4% (ranging from 6.3% to 29.0%) compared to locally trained models. Such generalization capability is particularly critical in hydrological forecasting, where data scarcity often impedes the application of sophisticated models. Overall, this study underscores the potential of ensemble machine learning models and transfer learning strategies to enhance the accuracy and reliability of tidal water level predictions, thereby offering robust support for flood management and ecological conservation efforts in coastal 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

Leveraging Transfer Learning for Improved Tidal Level Forecasting

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Yuan Liu,
  • Chao Wang,
  • Tongbi Tu

摘要

Accurate and rapid forecasting of nearshore water levels is crucial for managing estuarine and coastal ecosystems, where tidal dynamics, river discharge, and human activities interact in complex ways. This study employs AutoGluon, an automated ensemble machine learning framework, to predict tidal water levels in the Pearl River Delta and compares its performance against a Long Short-Term Memory (LSTM) model under diverse tidal conditions. Our results demonstrate that the proposed framework achieves predictive accuracy comparable to the LSTM baseline, as evaluated by the Nash-Sutcliffe Efficiency (NSE), while reducing model training time by nearly sevenfold. The potential of transfer learning for cross-regional prediction is further assessed to address data scarcity in ungauged areas. The transferred model attains NSE values exceeding 0.85 at unseen stations without prior local training, improving prediction accuracy by an average of 15.4% (ranging from 6.3% to 29.0%) compared to locally trained models. Such generalization capability is particularly critical in hydrological forecasting, where data scarcity often impedes the application of sophisticated models. Overall, this study underscores the potential of ensemble machine learning models and transfer learning strategies to enhance the accuracy and reliability of tidal water level predictions, thereby offering robust support for flood management and ecological conservation efforts in coastal regions.