Scholarlink

<p>In hydrology, streamflow forecasting with limited data poses significant challenges. Traditional Long Short-Term Memory (LSTM) networks struggle to adequately extract catchment information due to data scarcity, resulting in constrained prediction accuracy. This study proposes the TPE-selected LSTM-SVR blending-ensemble (LS) framework: utilizing Tree-structured Parzen Estimator (TPE) Bayesian optimization to screen multiple hyperparameter sets, constructing a collection of LSTM base learners with temporal-lag heterogeneity, and employing Support Vector Regression (SVR) as a meta-learner to fuse multi-physical perspective predictions. Furthermore, dedicated meta-learners are configured for each lead time, forming the LS-Based Step-Sequence (LSS) framework. Taking the Yiluo River Basin as a case study (with 3 years of training data), results from 100 runs show that for 1-, 3-, 5-, and 7-day lead times, the median Nash-Sutcliffe Efficiency (NSE) values are 0.89, 0.67, 0.63, and 0.64 for LS, and 0.88, 0.71, 0.66, and 0.66 for LSS, respectively—both outperforming LSTM-SS (0.80, 0.69, 0.63, 0.65) with superior stability (no NSE &lt; 0 outliers). Entropy analysis further reveals that LSS possesses the mechanistic advantages of both stability and independence. This study demonstrates that physically-guided ensemble diversity and step-specific design can effectively enhance streamflow forecasting accuracy in data-limited scenarios. Our research provides a feasible approach for streamflow forecasting studies in data-limited basins.</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

Blending-ensemble LSTM-SVR Modeling for Streamflow Prediction with Limited Data Series Based on Tree-structured Parzen Estimator Bayesian Optimization

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Jiyu Zhang,
  • Yufeng Wang,
  • Liangjie Zhao,
  • Bing Yan,
  • Shiqin Xu,
  • Zhe Yang,
  • La Zhuo

摘要

In hydrology, streamflow forecasting with limited data poses significant challenges. Traditional Long Short-Term Memory (LSTM) networks struggle to adequately extract catchment information due to data scarcity, resulting in constrained prediction accuracy. This study proposes the TPE-selected LSTM-SVR blending-ensemble (LS) framework: utilizing Tree-structured Parzen Estimator (TPE) Bayesian optimization to screen multiple hyperparameter sets, constructing a collection of LSTM base learners with temporal-lag heterogeneity, and employing Support Vector Regression (SVR) as a meta-learner to fuse multi-physical perspective predictions. Furthermore, dedicated meta-learners are configured for each lead time, forming the LS-Based Step-Sequence (LSS) framework. Taking the Yiluo River Basin as a case study (with 3 years of training data), results from 100 runs show that for 1-, 3-, 5-, and 7-day lead times, the median Nash-Sutcliffe Efficiency (NSE) values are 0.89, 0.67, 0.63, and 0.64 for LS, and 0.88, 0.71, 0.66, and 0.66 for LSS, respectively—both outperforming LSTM-SS (0.80, 0.69, 0.63, 0.65) with superior stability (no NSE < 0 outliers). Entropy analysis further reveals that LSS possesses the mechanistic advantages of both stability and independence. This study demonstrates that physically-guided ensemble diversity and step-specific design can effectively enhance streamflow forecasting accuracy in data-limited scenarios. Our research provides a feasible approach for streamflow forecasting studies in data-limited basins.