Scholarlink

<p>Runoff prediction is critical for water resource management. Multi-model integration presents multiple advantages over individual prediction models. Nevertheless, two major challenges remain: fixed or manually assigned weights cannot capture the simulation performance of each model, and runoff exhibits temporal heterogeneity, which requires time-varying weight allocation. This study proposes an optimisation method for weight allocation in combined prediction models based on a predefined objective function (WAOF) and identifies the Optimal Training set Time Window (OTTW) to address these issues. We first evaluated typical data-driven runoff models, then developed the WAOF framework and compared it with five alternative weighting methods, and finally discussed the effects of different training durations. Case studies were conducted at six hydrological stations in the Yangtze River Basin. The results show that Polynomial Regression (PR) and Multiple Linear Regression (MLR) models have robust applicability, whereas deep learning models suffer from poor stability, and no single model performs optimally at all stations. Under the optimal foresight period (OFP), the WAOF significantly outperforms the other methods (Diebold-Mariano test, <i>p</i> &lt; 0.01). Using 6–9 year of training data (OTTW), WAOF achieves a comparable accuracy to that of the full dataset. At Beibei Station, the differences in the MAE and QR are less than 0.5%, with an RMSE reduction of 1.3%. The proposed WAOF and OTTW provide reliable references for practical runoff prediction.</p> Graphical Abstract <p></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

Research on Multi Model Combination Prediction Strategy for Short term Runoff Prediction Based on Data Driven Approach

  • 期刊论文

  • 发表日期：

    2026年7月28日

查看全文
  • Zhongzheng He,
  • Jiahao Lu,
  • Yongqiang Wang,
  • Chao Wang,
  • Taixin Zhang,
  • Jun Guo,
  • Haimeng Guo,
  • Hui Qin

摘要

Runoff prediction is critical for water resource management. Multi-model integration presents multiple advantages over individual prediction models. Nevertheless, two major challenges remain: fixed or manually assigned weights cannot capture the simulation performance of each model, and runoff exhibits temporal heterogeneity, which requires time-varying weight allocation. This study proposes an optimisation method for weight allocation in combined prediction models based on a predefined objective function (WAOF) and identifies the Optimal Training set Time Window (OTTW) to address these issues. We first evaluated typical data-driven runoff models, then developed the WAOF framework and compared it with five alternative weighting methods, and finally discussed the effects of different training durations. Case studies were conducted at six hydrological stations in the Yangtze River Basin. The results show that Polynomial Regression (PR) and Multiple Linear Regression (MLR) models have robust applicability, whereas deep learning models suffer from poor stability, and no single model performs optimally at all stations. Under the optimal foresight period (OFP), the WAOF significantly outperforms the other methods (Diebold-Mariano test, p < 0.01). Using 6–9 year of training data (OTTW), WAOF achieves a comparable accuracy to that of the full dataset. At Beibei Station, the differences in the MAE and QR are less than 0.5%, with an RMSE reduction of 1.3%. The proposed WAOF and OTTW provide reliable references for practical runoff prediction.

Graphical Abstract