Accurate prediction of streamflow and understanding its influential features are crucial for ensuring the reliability of models used in managing water resources effectively at a basin level. This study conducts a comparative analysis of different models including univariate deep learning, process-based model (PBM), and a combination of PBM and deep learning, to enhance the accuracy of streamflow prediction. The study introduces the univariate N-BEATS model, renowned for its proficiency in analyzing single-variable time series, and a standard LSTM model to predict daily streamflow at three river gauge stations in the Ponnaiyar River Basin. Despite the N-BEATS univariate model showing promising accuracy during testing at Gummanur (R2 = 0.75), Vazhavachanur (R2 = 0.71), and Villupuram (R2 = 0.67) stations compared to the LSTM model, the study underscores the importance of considering the intricate relationships among multiple variables in the river basin. Ignoring these complexities could result in suboptimal predictive accuracy for the models. Therefore, the PBM SWAT model was established to predict daily streamflow, exhibiting lower accuracy during calibration at Gummanur (R2 = 0.38), Vazhavachanur (R2 = 0.32), and Villupuram (R2 = 0.26) stations. To enhance predictability while maintaining accuracy and reliability, this study proposed integrating influential features from calibrated SWAT-generated features using Pearson correlation analysis and the interpretable SHAP technique with XGBoost. Incorporating significant positive and negative impact features identified through SHAP analysis, the study developed coupled PBM and deep learning SWAT-N-BEATS and SWAT-LSTM, along with the SWAT-XGBoost model, to improve daily streamflow prediction. The SWAT-N-BEATS model exhibited superior accuracy during testing at Gummanur (R2 = 0.82), Vazhavachanur (R2 = 0.75), and Villupuram (R2 = 0.74) stations compared to the SWAT-LSTM, SWAT-XGBoost, univariate LSTM, N-BEATS and calibrated SWAT model. This research emphasizes the significance of comprehending the features that influence streamflow prediction, highlights the effectiveness of N-BEATS in streamflow prediction, and demonstrates the benefits of the coupled model compared to univariate streamflow prediction using deep learning techniques, especially in regions with limited data availability.

Enhancing daily streamflow prediction: A comparative analysis of univariate LSTM and N-BEATS models with coupled SWAT-LSTM and SWAT-N-BEATS models incorporating influential SWAT features

期刊论文
发表日期：

2025年5月17日

查看全文

R Yamini Priya,
R Manjula

摘要

Accurate prediction of streamflow and understanding its influential features are crucial for ensuring the reliability of models used in managing water resources effectively at a basin level. This study conducts a comparative analysis of different models including univariate deep learning, process-based model (PBM), and a combination of PBM and deep learning, to enhance the accuracy of streamflow prediction. The study introduces the univariate N-BEATS model, renowned for its proficiency in analyzing single-variable time series, and a standard LSTM model to predict daily streamflow at three river gauge stations in the Ponnaiyar River Basin. Despite the N-BEATS univariate model showing promising accuracy during testing at Gummanur (R² = 0.75), Vazhavachanur (R² = 0.71), and Villupuram (R² = 0.67) stations compared to the LSTM model, the study underscores the importance of considering the intricate relationships among multiple variables in the river basin. Ignoring these complexities could result in suboptimal predictive accuracy for the models. Therefore, the PBM SWAT model was established to predict daily streamflow, exhibiting lower accuracy during calibration at Gummanur (R² = 0.38), Vazhavachanur (R² = 0.32), and Villupuram (R² = 0.26) stations. To enhance predictability while maintaining accuracy and reliability, this study proposed integrating influential features from calibrated SWAT-generated features using Pearson correlation analysis and the interpretable SHAP technique with XGBoost. Incorporating significant positive and negative impact features identified through SHAP analysis, the study developed coupled PBM and deep learning SWAT-N-BEATS and SWAT-LSTM, along with the SWAT-XGBoost model, to improve daily streamflow prediction. The SWAT-N-BEATS model exhibited superior accuracy during testing at Gummanur (R² = 0.82), Vazhavachanur (R² = 0.75), and Villupuram (R² = 0.74) stations compared to the SWAT-LSTM, SWAT-XGBoost, univariate LSTM, N-BEATS and calibrated SWAT model. This research emphasizes the significance of comprehending the features that influence streamflow prediction, highlights the effectiveness of N-BEATS in streamflow prediction, and demonstrates the benefits of the coupled model compared to univariate streamflow prediction using deep learning techniques, especially in regions with limited data availability.