Machine-learning- and deep-learning-based streamflow prediction in a hilly catchment for future scenarios using CMIP6 GCM data

Abstract

Abstract. The alteration in river flow patterns, particularly those that originate in the Himalaya, has been caused by the increased temperature and rainfall variability brought on by climate change. Due to the impending intensification of extreme climate events, as predicted by the Intergovernmental Panel on Climate Change (IPCC) in its Sixth Assessment Report, it is more essential than ever to predict changes in streamflow for future periods. Despite the fact that some research has utilised machine-learning- and deep-learning-based models to predict streamflow patterns in response to climate change, very few studies have been undertaken for a mountainous catchment, with the number of studies for the western Himalaya being minimal. This study investigates the capability of five different machine learning (ML) models and one deep learning (DL) model, namely the Gaussian linear regression model (GLM), Gaussian generalised additive model (GAM), multivariate adaptive regression splines (MARSs), artificial neural network (ANN), random forest (RF), and 1D convolutional neural network (1D-CNN), in streamflow prediction over the Sutlej River basin in the western Himalaya during the periods 2041–2070 (2050s) and 2071–2100 (2080s). Bias-corrected data downscaled at a grid resolution of 0.25∘ × 0.25∘ from six general circulation models (GCMs) of the Coupled Model Intercomparison Project Phase 6 GCM framework under two greenhouse gas (GHG) trajectories (SSP245 and SSP585) were used for this purpose. Four different rainfall scenarios (R0, R1, R2, and R3) were applied to the models trained with daily data (1979–2009) at Kasol (the outlet of the basin) in order to better understand how catchment size and the geo-hydromorphological aspects of the basin affect runoff. The predictive power of each model was assessed using six statistical measures, i.e. the coefficient of determination (R2), the ratio of the root mean square error to the standard deviation of the measured data (RSR), the mean absolute error (MAE), the Kling–Gupta efficiency (KGE), the Nash–Sutcliffe efficiency (NSE), and the percent bias (PBIAS). The RF model with rainfall scenario R3, which outperformed other models during the training (R2 = 0.90; RSR = 0.32; KGE = 0.87; NSE = 0.87; PBIAS = 0.03) and testing (R2 = 0.78; RSR = 0.47; KGE = 0.82; NSE = 0.71; PBIAS = −0.31) period, therefore was chosen to simulate streamflow in the Sutlej River in the 2050s and 2080s under the SSP245 and SSP585 scenarios. Bias correction was further applied to the projected daily streamflow in order to generate a reliable times series of the discharge. The mean ensemble of the model results shows that the mean annual streamflow of the Sutlej River is expected to rise between 2050s and 2080s by 0.79 % to 1.43 % for SSP585 and by 0.87 % to 1.10 % for SSP245. In addition, streamflow will increase during the monsoon (9.70 % to 11.41 % and 11.64 % to 12.70 %) in the 2050s and 2080s under both emission scenarios, but it will decrease during the pre-monsoon (−10.36 % to −6.12 % and −10.0 % to −9.13 %), post-monsoon (−1.23 % to −0.22 % and −5.59 % to −2.83 %), and during the winter (−21.87 % to −21.52 % and −21.87 % to −21.11 %). This variability in streamflow is highly correlated with the pattern of precipitation and temperature predicted by CMIP6 GCMs for future emission scenarios and with physical processes operating within the catchment. Predicted declines in the Sutlej River streamflow over the pre-monsoon (April to June) and winter (December to March) seasons might have a significant impact on agriculture downstream of the river, which is already having problems due to water restrictions at this time of year. The present study will therefore assist in strategy planning to ensure the sustainable use of water resources downstream by acquiring knowledge of the nature and causes of unpredictable streamflow patterns.