[A-HW21] Hydrological modelling to support water resources management and engineering designs

[AHW21-03] Optimizing the accuracy of peak flow prediction by using the SWAT model with machine learning techniques in the Beas River basin, India

*Saran raaj¹, Vivek Gupta¹, Vishal Singh², Dericks Shukla¹ (1.School of Civil and Environmental Engineering, Indian Institue of Technology Mandi, India, 2.Water Resources System Division, National Institue of Hydrology Roorkee, India)

Keywords:SWAT, Machine Learning, Flood, Peak flow

Reliable extreme streamflow simulation has been of interest to scientists and resource managers during the last century. Rainfall-runoff models, which are normally used to estimate stream flow accurately at the river basin scale often lacks the accuracy in simulating the extreme flow values. Soil and Water Assessment Tool (SWAT) is one amongst these rainfall-runoff models, which has been used in predicting the stream flow values. In this study, machine learning based regression models have been used to predict accurately the stream flow with a focus on extreme event simulation in combination with uncalibrated SWAT (uSWAT-ML) outputs. Eight ML regression models including linear regression (LR), multi-layer perceptron (MLP), extreme gradient boosting (XGBoost), light gradient-boosting machine (LGBM), kernel ridge (KR), elastic net (EN), Bayesian ridge (BR), and gradient boosting (GB) have been used to perform and analyze the model. For calculating the efficiency of the model, the coefficient of determination (R²), Nash-Sutcliffe efficiency (NSE) and root mean square error (RMSE) were considered. The results indicated that physics aided ML models performed well as compared to calibrated SWAT model. R² value of 0.73, NSE value of 0.72 and RMSE value of 276.92 m³/s shows good performance of calibrated SWAT model. Whereas in uSWAT-ML model, except for the LR model all other model performed better than the calibrated SWAT model. Among the eight uSWAT-ML model, EN and BR have obtained better results with R² value of 0.89 and 0.89, NSE value of 0.87 and 0.87, and RMSE value of 158.31 m³/s and 159.48 m³/s. Predicting the peak flow values have also been done, and uSWAT-ML have predicted better with EN and BR results with R² value 0.71 and 0.71. The approach of predicting stream flow value in combination with uncalibrated SWAT and machine learning techniques was found to enhance the reliability of the results as compared to tradition SWAT calibration method.