Scholarlink

<p>Monitoring water release volumes (WRVs) from dams is essential for sustainable water management, as WRVs directly affect water allocation planning and drought mitigation strategies. The aim of this study is to develop accurate forecasting models to support dam operation and decision-making. To this end, a 15-year dataset (2002–2017) from the Hammam Debagh (HD) dam in Algeria was used to compare five supervised learning algorithms: Multilayer Perceptron Neural Network (MLPNN), Random Forest (RFR), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and XGBoost. The Gamma Test (GT) was used to identify the seven most suitable input combinations, and the dataset was divided into 70% for training and 30% for testing. Unlike previous studies, which have mainly focused on predicting inflow, storage or evaporation, this work targets the operational forecasting of WRV a critical but rarely modelled variable in reservoir management. The study’s novelty lies in performing a cross-family comparison of ensemble, feedforward, and recurrent architectures on a long-term dataset under a consistent validation framework. The results show that all the models achieved strong predictive performance (R and NSE close to 1). However, LSTM-based models outperformed the others in capturing temporal dependencies. The optimal configuration (LSTM6) achieved <i>R</i> = 0.976, NSE = 0.951, KGE = 0.972, MAE = 0.010 and RMSE = 0.025, demonstrating high precision and robustness. These findings confirm the suitability of LSTM architectures for operational WRV forecasting, as well as their potential for integration into proactive reservoir management systems to support sustainable water resource planning.</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

Enhancing Forecasting of Water Release Volumes from Dams: A Comparative Study of Supervised Learning Models

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Mehdi Amiour,
  • Toufik Bouziane,
  • Noureddine Daif,
  • Djamel Bengora

摘要

Monitoring water release volumes (WRVs) from dams is essential for sustainable water management, as WRVs directly affect water allocation planning and drought mitigation strategies. The aim of this study is to develop accurate forecasting models to support dam operation and decision-making. To this end, a 15-year dataset (2002–2017) from the Hammam Debagh (HD) dam in Algeria was used to compare five supervised learning algorithms: Multilayer Perceptron Neural Network (MLPNN), Random Forest (RFR), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and XGBoost. The Gamma Test (GT) was used to identify the seven most suitable input combinations, and the dataset was divided into 70% for training and 30% for testing. Unlike previous studies, which have mainly focused on predicting inflow, storage or evaporation, this work targets the operational forecasting of WRV a critical but rarely modelled variable in reservoir management. The study’s novelty lies in performing a cross-family comparison of ensemble, feedforward, and recurrent architectures on a long-term dataset under a consistent validation framework. The results show that all the models achieved strong predictive performance (R and NSE close to 1). However, LSTM-based models outperformed the others in capturing temporal dependencies. The optimal configuration (LSTM6) achieved R = 0.976, NSE = 0.951, KGE = 0.972, MAE = 0.010 and RMSE = 0.025, demonstrating high precision and robustness. These findings confirm the suitability of LSTM architectures for operational WRV forecasting, as well as their potential for integration into proactive reservoir management systems to support sustainable water resource planning.