Machine learning and bias corrected CMIP6 for improved streamflow prediction in the data-scarce Kulfo river, Ethiopia

Climate change is increasingly affecting hydrological processes and water availability, particularly in data-scarce regions such as Ethiopia’s Rift Valley basins. However, few studies have systematically integrated climate model bias correction, machine learning-based prediction, uncertainty quantification, and non-stationary flood assessment within a unified framework for climate-resilient water resources management. This study developed an integrated climate–hydrology–machine learning framework using bias-corrected Coupled Model Intercomparison Project Phase 6 (CMIP6) projections to evaluate future dynamics in the Kulfo watershed, southern Ethiopia. Sixteen CMIP6 models under SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios were assessed, and statistical and machine learning-based bias-correction approaches were compared. Random Forest (RF) demonstrated the best performance for both bias correction and prediction, reducing precipitation errors by up to 92% and achieving Nash–Sutcliffe Efficiency (NSE) values of 0.999 during calibration and 0.806 during validation. Future projections indicate increasing temperatures and a progressive decline in annual by approximately 8.5, 13.9, and 20.7% under SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios, respectively, by 2100. These changes are associated with increased evaporative demand and altered seasonal rainfall patterns, leading to reduced summer discharge and shifts in hydrological seasonality. Conformal prediction revealed increasing uncertainty in future projections, particularly under the high-emission SSP5-8.5 scenario, while non-stationary flood frequency analysis indicated significant changes in future flood characteristics. The proposed framework improves understanding of climate-driven hydrological changes and provides a robust approach for prediction and uncertainty assessment in data-scarce catchments. The findings support sustainable water resources management, climate adaptation, irrigation planning, and flood risk reduction, contributing to Sustainable Development Goals (SDGs) 2 (Zero Hunger), 6 (Clean Water and Sanitation), 13 (Climate Action), and 15 (Life on Land). Better-performing CMIP6 models: CNRM-ESM2-1 for precipitation and minimum temperature, and CNRM-CM6-1 for maximum temperature after bias correction Random Forest bias correction reduced precipitation RMSE by 92% Comprehensive comparison of 10 machine learning algorithms revealed Random Forest achieved strong performance (NSE = 0.999 training; 0.806 validation) The superior performance of Random Forest over LSTM and deep learning methods in this data-scarce catchment demonstrates that simpler ensemble methods can be more appropriate for limited training data Climate projections show rising temperatures and declining streamflowunder all SSP scenarios Conformal prediction quantified uncertainty using 95% prediction intervals.