Scholarlink

<p>Reliable streamflow prediction is essential for flood management, hydropower operation, irrigation scheduling, and climate change adaptation, particularly in large, transboundary, data-scarce river basins. While many studies apply deep learning (DL) approaches using historical discharge data, relatively few explore climate-only DL frameworks for daily streamflow prediction. This study develops a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model to simulate daily streamflow in the Brahmaputra River Basin, one of the world’s largest and most climate-sensitive transboundary systems. Unlike most DL-based approaches, the model relies exclusively on universally available climatic predictors—precipitation and temperature—from the W5E5 reanalysis dataset. Observed discharge data from the Bangladesh Water Development Board (BWDB) were used solely as target outputs and not as input features. The dataset was divided into training (1985–2002) and testing (2003–2014) periods, and lag analysis identified a 15-day input window as optimal for capturing rainfall–runoff dependencies. Results show that the CNN–LSTM consistently outperformed standalone CNN, LSTM, and Multilayer Perceptron (MLP) models, achieving Nash–Sutcliffe Efficiency (NSE) values of 0.92 during training and 0.86 during testing. Rainfall emerged as the dominant driver of streamflow variability, while temperature enhanced predictive skill during low-flow and transitional periods. Benchmarking against a calibrated SWAT model under identical climatic forcing showed that CNN–LSTM achieved comparable or superior performance, particularly for low-flow conditions, high-flow extremes, and peak-flow timing. Overall, the findings demonstrate the potential of climate-driven DL frameworks as robust tools for daily streamflow prediction in data-scarce basins, supporting flood forecasting, water resources management, and climate adaptation under increasing hydroclimatic uncertainty.</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 Streamflow Prediction Using a Hybrid CNN-LSTM Model: Integrating Climate-Driven Spatio-Temporal Features in a Deep Learning Framework

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Shadman Shahariar,
  • Hasan Zobeyer,
  • Nasreen Jahan,
  • Mostafa Ali

摘要

Reliable streamflow prediction is essential for flood management, hydropower operation, irrigation scheduling, and climate change adaptation, particularly in large, transboundary, data-scarce river basins. While many studies apply deep learning (DL) approaches using historical discharge data, relatively few explore climate-only DL frameworks for daily streamflow prediction. This study develops a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model to simulate daily streamflow in the Brahmaputra River Basin, one of the world’s largest and most climate-sensitive transboundary systems. Unlike most DL-based approaches, the model relies exclusively on universally available climatic predictors—precipitation and temperature—from the W5E5 reanalysis dataset. Observed discharge data from the Bangladesh Water Development Board (BWDB) were used solely as target outputs and not as input features. The dataset was divided into training (1985–2002) and testing (2003–2014) periods, and lag analysis identified a 15-day input window as optimal for capturing rainfall–runoff dependencies. Results show that the CNN–LSTM consistently outperformed standalone CNN, LSTM, and Multilayer Perceptron (MLP) models, achieving Nash–Sutcliffe Efficiency (NSE) values of 0.92 during training and 0.86 during testing. Rainfall emerged as the dominant driver of streamflow variability, while temperature enhanced predictive skill during low-flow and transitional periods. Benchmarking against a calibrated SWAT model under identical climatic forcing showed that CNN–LSTM achieved comparable or superior performance, particularly for low-flow conditions, high-flow extremes, and peak-flow timing. Overall, the findings demonstrate the potential of climate-driven DL frameworks as robust tools for daily streamflow prediction in data-scarce basins, supporting flood forecasting, water resources management, and climate adaptation under increasing hydroclimatic uncertainty.