Scholarlink

<p>Accurate daily inflow forecasting is essential for extreme climate adaptation, flood control, and hydropower optimization. However, rapid climate change demands frequent model reconstruction and parameter recalibration, which undermines daily streamflow prediction accuracy, while the poor interpretability of deep learning models applied to this domain poses an additional challenge. To address these issues, a one-day-ahead inflow forecasting framework of cross-attention hybrid random forest and long short-term memory (CA-RF-LSTM) integrated with shapley additive explanations (SHAP) is proposed. First, a multifactor input model coupling observed and meteorological data with lag times is constructed to enhance model adaptability under changing hydroclimatic conditions and reduce the dependence on repeated model restructuring and parameter recalibration. Second, a cross-attention fusion mechanism is introduced to adaptively integrate nonlinear spatial relationships extracted by the RF and temporal dependencies learned by the LSTM, enabling dynamic interactions between heterogeneous feature representations and improving the predictive accuracy beyond conventional feature concatenation strategies. Finally, SHAP is employed to quantify feature contributions, optimize input combinations, and provide transparent interpretations of model behavior, thereby enhancing model explainability. The case study for one-day-ahead forecasting indicates that the proposed framework achieves overall superior predictive performance relative to the benchmark methods according to six evaluation metrics and the Wilcoxon signed-rank test result, confirming the effectiveness of the proposed cross-attention fusion strategy and SHAP-guided input optimization.</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

Interpretable Daily Inflow Forecasting Model via Cross-Attention Hybrid Neural Networks Integrated with SHAP Analysis

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Xiaoyan Wu,
  • Hongye Zhao,
  • Huaying Su,
  • Huan Wang,
  • Jiang Xiong,
  • Shengli Liao

摘要

Accurate daily inflow forecasting is essential for extreme climate adaptation, flood control, and hydropower optimization. However, rapid climate change demands frequent model reconstruction and parameter recalibration, which undermines daily streamflow prediction accuracy, while the poor interpretability of deep learning models applied to this domain poses an additional challenge. To address these issues, a one-day-ahead inflow forecasting framework of cross-attention hybrid random forest and long short-term memory (CA-RF-LSTM) integrated with shapley additive explanations (SHAP) is proposed. First, a multifactor input model coupling observed and meteorological data with lag times is constructed to enhance model adaptability under changing hydroclimatic conditions and reduce the dependence on repeated model restructuring and parameter recalibration. Second, a cross-attention fusion mechanism is introduced to adaptively integrate nonlinear spatial relationships extracted by the RF and temporal dependencies learned by the LSTM, enabling dynamic interactions between heterogeneous feature representations and improving the predictive accuracy beyond conventional feature concatenation strategies. Finally, SHAP is employed to quantify feature contributions, optimize input combinations, and provide transparent interpretations of model behavior, thereby enhancing model explainability. The case study for one-day-ahead forecasting indicates that the proposed framework achieves overall superior predictive performance relative to the benchmark methods according to six evaluation metrics and the Wilcoxon signed-rank test result, confirming the effectiveness of the proposed cross-attention fusion strategy and SHAP-guided input optimization.