Application of coupling machine learning techniques and linear Bias scaling for optimizing 10-daily flow simulations, Swat River Basin

Abstract

Accurate hydrological simulations comply with the water (sixth) Sustainable Development Goals (SDGs). The study investigates the utility of artificial neural networks and support vector regressors, as well as the post-simulation bias treatment of these simulations in the Swat River basin, Pakistan. For this, climate variables were lag adjusted for the first time, then cross-correlated with the flow to identify the most associative delay time. In sensitivity analysis, seven combinations were selected as input with suitable hyperparameters. For SVR, grid search cross-validation was used for determining the optimal set of hyperparameters, while for ANN, neurons and hidden layers were optimized by trial and error. We ran the model by using optimized hyperparameter configurations and input combinations. In comparison to SVRs (Root mean square error (RMSE) 34.2; mean absolute error (MAE) 3.0; CC 0.91) values, respectively, ANN fits the observations better than SVR with (RMSE 11.9; MAE 1.14; CC 0.99). Linear bias-corrected simulations greatly improved ANN performance (RMSE 3.98; MAE 0.625; CC 0.99), while the improvement was slight in the case of SVR (RMSE 35; MAE 0.58; CC 0.92). On a seasonal scale, bias-corrected simulations remedy the low- and high-flow seasonal discrepancies. Flow duration analysis results reveal deviation at low- and high-flow conditions by models, which were then reconciled by applying bias corrections.