Scholarlink

<p>Accurate runoff prediction is crucial for water resource management and flood disaster prevention. However, the non-linear and non-stationary characteristics of runoff series make precise forecasting a complex task. To address this challenge, this study proposes an integrated ensemble model, termed SGMD-SE-BKA-BiTCN-EC, for medium- to long-term runoff prediction through the integration of symplectic geometric mode decomposition (SGMD), sample entropy (SE), the black-winged kite algorithm (BKA), a bidirectional temporal convolutional network (BiTCN), and error correction (EC). The model framework operates as follows: First, the SGMD decomposes historical runoff data into multiple symplectic geometric components (SGCs). These components are then reconstructed using SE to generate multiple subsequences, thereby improving input data quality and enhancing model efficiency. Second, the BKA is employed to optimize the hyperparameters of the BiTCN model. The hybrid SGMD-SE-BKA-BiTCN model subsequently predicts each subsequence, and the predicted results are aggregated to form the preliminary prediction. Finally, an EC module corrects residual errors in the preliminary prediction, and these adjustments are integrated with the initial results to produce the final monthly runoff prediction values. The proposed model was validated across two heterogeneous basins (namely, the arid Heihe River and the humid Lancang River basins) and demonstrated exceptional accuracy, achieving a mean Kling-Gupta efficiency (KGE) of 0.979 and a Nash–Sutcliffe efficiency coefficient (NSEC) of 0.983. It outperformed 11 benchmark models, reducing extreme prediction errors by up to 87.4% through error correction while maintaining high computational efficiency (training time &lt; 180 s). Consequently, the proposed model provides a robust analytical tool for hydrological forecasting under climate change, offering critical insights for adaptive water allocation and extreme-event preparedness across diverse geographical 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

A Hybrid Deep Learning Framework Integrating Multi-Strategy Optimization and Error Correction for Ensemble Monthly Runoff Prediction

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Wen-chuan Wang,
  • Miao Gu,
  • Ming-wei Ma,
  • Hong-fei Zang,
  • Qiang Ma,
  • Dong-mei Xu

摘要

Accurate runoff prediction is crucial for water resource management and flood disaster prevention. However, the non-linear and non-stationary characteristics of runoff series make precise forecasting a complex task. To address this challenge, this study proposes an integrated ensemble model, termed SGMD-SE-BKA-BiTCN-EC, for medium- to long-term runoff prediction through the integration of symplectic geometric mode decomposition (SGMD), sample entropy (SE), the black-winged kite algorithm (BKA), a bidirectional temporal convolutional network (BiTCN), and error correction (EC). The model framework operates as follows: First, the SGMD decomposes historical runoff data into multiple symplectic geometric components (SGCs). These components are then reconstructed using SE to generate multiple subsequences, thereby improving input data quality and enhancing model efficiency. Second, the BKA is employed to optimize the hyperparameters of the BiTCN model. The hybrid SGMD-SE-BKA-BiTCN model subsequently predicts each subsequence, and the predicted results are aggregated to form the preliminary prediction. Finally, an EC module corrects residual errors in the preliminary prediction, and these adjustments are integrated with the initial results to produce the final monthly runoff prediction values. The proposed model was validated across two heterogeneous basins (namely, the arid Heihe River and the humid Lancang River basins) and demonstrated exceptional accuracy, achieving a mean Kling-Gupta efficiency (KGE) of 0.979 and a Nash–Sutcliffe efficiency coefficient (NSEC) of 0.983. It outperformed 11 benchmark models, reducing extreme prediction errors by up to 87.4% through error correction while maintaining high computational efficiency (training time < 180 s). Consequently, the proposed model provides a robust analytical tool for hydrological forecasting under climate change, offering critical insights for adaptive water allocation and extreme-event preparedness across diverse geographical regions.