Evaluation of machine learning models for streamflow projections under different greenhouse gas emission scenarios

With the increasing demand for precise water management under climate change, hydrological prediction has become more important than ever. In this context, machine learning emerges as a powerful tool to improve forecasting performance. This study evaluates the performance of LightGBM, Random Forest (RF) and Ridge Regression (RR) models for streamflow prediction and climate change impact assessment. Using historical streamflow data from 1984 to 2014, model accuracy was assessed with key performance metrics. LightGBM outperformed RF and RR, achieving lower RMSE (3.046 m³/s), MAE (2.145 m³/s), RSR (0.599), and higher NSE (0.64) and KGE (0.683), demonstrating greater accuracy and robustness due to its gradient-boosting framework. Based on these results, LightGBM was used to project future streamflow under CMIP6 climate scenarios with the NorESM2 model. Projections show increasing streamflow trends under SSP3-7.0 (3.436%) and SSP5-8.5 (4.049%), a decrease under SSP1-2.6, and a stable trend under SSP2-4.5, according to historical streamflow data. The projected future streamflow under these climate scenarios provides valuable insights for water resource management, highlighting the importance of using advanced machine learning techniques like LightGBM for hydrological forecasting in the context of climate change. • LightGBM and Random Forest have been compared about performance of projection with each other. • LightGBM has been used to project streamflow in the future under various emission scenarios. • The findings show that the LightGBM model significantly improved performance prediction accuracy.