Abstract Accurate streamflow prediction is crucial for disaster prevention, mitigation, and water resource management. While data-driven approaches surpass the traditional process-based models, they are often questioned for their physical interpretability. Hybrid modelling frameworks that integrate hydrological process knowledge with data-driven techniques have emerged as a promising strategy to enhance streamflow prediction accuracy while maintaining physical interpretability. This study explores the potential of hybrid models that integrate the physically based MIKE NAM model with advanced machine learning techniques such as Long Short Term Memory (LSTM). Two hybrid configurations were developed: one using MIKE NAM-simulated streamflow as input to the LSTM, and another using both MIKE NAM-simulated streamflow and individual NAM components (overland flow, baseflow, interflow, and evapotranspiration) to enhance residual learning. The models are applied to predict 1-day streamflow in the Meenachil River Basin, India, and the performance of each model is evaluated using MAE, RMSE, R2, NSE, and KGE metrics across full-period and disaggregated flow regimes such as low, medium, and high flows. Results indicate that hybrid models significantly outperform both standalone MIKE NAM and LSTM models, even in diverse hydrological conditions (Kling Gupta Efficiency, KGEMIKE = 0.62, KGELSTM = 0.75, KGEHYBRID_I = 0.94, KGEHYBRID_2 = 0.93). This work demonstrates the potential of hybrid frameworks in bridging the gap between physically based understanding and data-driven accuracy, providing a reliable solution for operational forecasting and water resource planning in data-scarce and hydrologically complex environments. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> Hybrid NAM-LSTM models enhance streamflow prediction (KGE=0.93), outperforming standalone MIKE NAM (KGE = 0.62) and LSTM model (KGE = 0.75). </ItemContent> <ItemContent> Integration of NAM components enhances residual learning, improving sensitivity to hydrological processes. </ItemContent> <ItemContent> Hybrid models demonstrate strong potential for operational forecasting applications as demonstrated by enhanced extreme predictions. </ItemContent> </UnorderedList>

Bridging process-based hydrology and deep learning: exploring hybrid MIKE NAM–LSTM techniques for accurate streamflow prediction in a humid tropical river basin in India

期刊论文
发表日期：

2026年4月13日

查看全文

Divya Chandran,
Haleema Wardah,
N R Chithra

摘要

Abstract

Accurate streamflow prediction is crucial for disaster prevention, mitigation, and water resource management. While data-driven approaches surpass the traditional process-based models, they are often questioned for their physical interpretability. Hybrid modelling frameworks that integrate hydrological process knowledge with data-driven techniques have emerged as a promising strategy to enhance streamflow prediction accuracy while maintaining physical interpretability. This study explores the potential of hybrid models that integrate the physically based MIKE NAM model with advanced machine learning techniques such as Long Short Term Memory (LSTM). Two hybrid configurations were developed: one using MIKE NAM-simulated streamflow as input to the LSTM, and another using both MIKE NAM-simulated streamflow and individual NAM components (overland flow, baseflow, interflow, and evapotranspiration) to enhance residual learning. The models are applied to predict 1-day streamflow in the Meenachil River Basin, India, and the performance of each model is evaluated using MAE, RMSE, R², NSE, and KGE metrics across full-period and disaggregated flow regimes such as low, medium, and high flows. Results indicate that hybrid models significantly outperform both standalone MIKE NAM and LSTM models, even in diverse hydrological conditions (Kling Gupta Efficiency, KGE_MIKE = 0.62, KGE_LSTM = 0.75, KGE_{HYBRID_I} = 0.94, KGE_{HYBRID_2} = 0.93). This work demonstrates the potential of hybrid frameworks in bridging the gap between physically based understanding and data-driven accuracy, providing a reliable solution for operational forecasting and water resource planning in data-scarce and hydrologically complex environments.

Research highlights

Hybrid NAM-LSTM models enhance streamflow prediction (KGE=0.93), outperforming standalone MIKE NAM (KGE = 0.62) and LSTM model (KGE = 0.75).

Integration of NAM components enhances residual learning, improving sensitivity to hydrological processes.

Hybrid models demonstrate strong potential for operational forecasting applications as demonstrated by enhanced extreme predictions.