Scholarlink

<p>Accurate modeling of reservoir inflow is critical for efficient reservoir operations, optimal water resource management, and reliable irrigation supply. Previous research has primarily focused on short-term prediction, particularly during the flood season, while the performance during the dry season has been relatively understudied. Therefore, this study aims to propose three new residual-error-based ensemble learning (REL) models to improve medium-range reservoir-inflow forecasting both in flood and dry seasons. The novelty of the proposed models was primary based on the combination of base and meta-learners, the multiple output framework, and the residual-error corrections. To explore and achieve the suitable base and meta-learners, ten diverse data-driven (DD) models were firstly utilized at distinct reservoirs over a lead time of 1–10&#xa0;days. To further examine the reservoir inflow predictability of the proposed REL models, the extensive investigations of the model performance analysis were conducted. The contribution of feature importance on the predictions during flood and dry seasons was also conducted using the SHapley additive explanation method. Results indicated that the long short-term memory (LSTM), categorical gradient boosting regression (CGBR), light gradient boosting machine regression (LGBMR), and random forest regression (RFR) models were found to be the base learners (i.e., optimal models) for Liyutan, Zengwen, Agongdian, and Mudan reservoirs. The application of the three proposed REL models to the other Shihmen reservoir resulted in the achievement of reliable and satisfactory robustness and applicability, as illustrated by the lead-time-averaged generalization ability performance, which ranged from 0.9 to 1.1. It can be concluded that the proposed three REL models achieved the superior ten-day predictability performance of reservoir inflow compared to the other four base learners. In consideration of all five reservoir sites, the proposed REL3 model yielded the entire-reservoirs-averaged improvements in terms of lead-time correlation coefficient of approximately 16.9%, 7.2%, 11.2% and 10.7% in comparison to the LSTM, CGBR, LGBMR, and RFR models. These results offer the valuable insights into reservoir flood management and early warning systems.</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

Enhanced medium-range prediction for reservoir inflows using residual-error-based ensemble learning models

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Wen-Dar Guo,
  • Wei-Bo Chen,
  • Chih-Hsin Chang

摘要

Accurate modeling of reservoir inflow is critical for efficient reservoir operations, optimal water resource management, and reliable irrigation supply. Previous research has primarily focused on short-term prediction, particularly during the flood season, while the performance during the dry season has been relatively understudied. Therefore, this study aims to propose three new residual-error-based ensemble learning (REL) models to improve medium-range reservoir-inflow forecasting both in flood and dry seasons. The novelty of the proposed models was primary based on the combination of base and meta-learners, the multiple output framework, and the residual-error corrections. To explore and achieve the suitable base and meta-learners, ten diverse data-driven (DD) models were firstly utilized at distinct reservoirs over a lead time of 1–10 days. To further examine the reservoir inflow predictability of the proposed REL models, the extensive investigations of the model performance analysis were conducted. The contribution of feature importance on the predictions during flood and dry seasons was also conducted using the SHapley additive explanation method. Results indicated that the long short-term memory (LSTM), categorical gradient boosting regression (CGBR), light gradient boosting machine regression (LGBMR), and random forest regression (RFR) models were found to be the base learners (i.e., optimal models) for Liyutan, Zengwen, Agongdian, and Mudan reservoirs. The application of the three proposed REL models to the other Shihmen reservoir resulted in the achievement of reliable and satisfactory robustness and applicability, as illustrated by the lead-time-averaged generalization ability performance, which ranged from 0.9 to 1.1. It can be concluded that the proposed three REL models achieved the superior ten-day predictability performance of reservoir inflow compared to the other four base learners. In consideration of all five reservoir sites, the proposed REL3 model yielded the entire-reservoirs-averaged improvements in terms of lead-time correlation coefficient of approximately 16.9%, 7.2%, 11.2% and 10.7% in comparison to the LSTM, CGBR, LGBMR, and RFR models. These results offer the valuable insights into reservoir flood management and early warning systems.