Scholarlink

<p>Faced with the critical issues of low prediction accuracy and difficulty in feature identification-making under the dual influences of climate variability and human activities in runoff prediction, this study innovatively proposes a runoff prediction model based on endogenous-exogenous feature fusion. The model synchronously analyzes the high-frequency fluctuation characteristics and low-frequency trend components of runoff sequences, integrates exogenous environmental factors such as precipitation and temperature, and constructs an Endogenous-Exogenous Feature Fusion Module (EFFM) to investigate the causes of hydrological elements. On this basis, a Temporal Context Fusion Unit (TCFU) is designed, using Long Short-Term Memory (LSTM) networks to capture the daily-scale runoff evolution patterns while introducing an Efficient Additive Attention (EAA) mechanism to enhance the extraction of key hydrological event features, significantly improving the expression capability of temporal features. A hybrid model called Endo-Exo Temporal Transformer (ETT) is constructed by cascading EFFM and TCFU modules onto the Transformer model. The Monte Carlo random sampling method is employed to establish multi-confidence prediction intervals, thereby enhancing the support for risk decision-making while maintaining prediction accuracy. Additionally, a systematic analysis of the contribution rates of different driving factors is conducted to deepen the understanding of hydrological process mechanisms. Experimental results from predicting daily runoff at four typical climate-morphological condition hydrological stations demonstrate that the ETT model significantly outperforms seven benchmark models, including Transformer and LSTM. For instance, at the Crystal River station, the model achieved Nash–Sutcliffe Efficiency (NSE) of 0.997 and Kling-Gupta Efficiency coefficient of 0.979, with peak flow prediction errors reduced by 68.3% compared to the average of the comparative group. The study shows that the ETT model can effectively address complex environmental stress, improving the accuracy and stability of runoff prediction.</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

Monte-Carlo-assisted endo-exo temporal transformer for high-confidence interval forecasting of daily runoff

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Xiao-xue Hu,
  • Dong-mei Xu,
  • Wen-chuan Wang,
  • Jun Wang,
  • Zong Li

摘要

Faced with the critical issues of low prediction accuracy and difficulty in feature identification-making under the dual influences of climate variability and human activities in runoff prediction, this study innovatively proposes a runoff prediction model based on endogenous-exogenous feature fusion. The model synchronously analyzes the high-frequency fluctuation characteristics and low-frequency trend components of runoff sequences, integrates exogenous environmental factors such as precipitation and temperature, and constructs an Endogenous-Exogenous Feature Fusion Module (EFFM) to investigate the causes of hydrological elements. On this basis, a Temporal Context Fusion Unit (TCFU) is designed, using Long Short-Term Memory (LSTM) networks to capture the daily-scale runoff evolution patterns while introducing an Efficient Additive Attention (EAA) mechanism to enhance the extraction of key hydrological event features, significantly improving the expression capability of temporal features. A hybrid model called Endo-Exo Temporal Transformer (ETT) is constructed by cascading EFFM and TCFU modules onto the Transformer model. The Monte Carlo random sampling method is employed to establish multi-confidence prediction intervals, thereby enhancing the support for risk decision-making while maintaining prediction accuracy. Additionally, a systematic analysis of the contribution rates of different driving factors is conducted to deepen the understanding of hydrological process mechanisms. Experimental results from predicting daily runoff at four typical climate-morphological condition hydrological stations demonstrate that the ETT model significantly outperforms seven benchmark models, including Transformer and LSTM. For instance, at the Crystal River station, the model achieved Nash–Sutcliffe Efficiency (NSE) of 0.997 and Kling-Gupta Efficiency coefficient of 0.979, with peak flow prediction errors reduced by 68.3% compared to the average of the comparative group. The study shows that the ETT model can effectively address complex environmental stress, improving the accuracy and stability of runoff prediction.