Enhancing climate projections via machine learning: Multi-model ensemble of precipitation and temperature in the source region of the Yellow River

Abstract

Accurately estimating future variations in precipitation and temperature in the Source Region of the Yellow River (SRYR) is critical, given its key role as the water conservation area of the Yellow River basin in China. To enhance the accuracy of regional climate projections, this study proposes a machine learning-enhanced ensemble framework comprising global climate model (GCM) selection, bias correction, and multi-model ensemble, effectively combining physical climate modeling with data-driven methods. Specifically, a rank score method was used to evaluate the performance of 22 GCMs from the Coupled Model Intercomparison Project Phase 6 (CMIP6) in reproducing precipitation and temperature over the SRYR. Based on this assessment, six top-performing GCMs were selected and subsequently bias-corrected. To determine the most effective ensemble strategy, four multi-model ensemble approaches—including weighted averaging (WA), random forest (RF), feedforward neural network (FNN), and long short-term memory (LSTM)—were employed to integrate the bias-corrected outputs from the selected models over the historical period. All ensemble approaches outperformed individual GCMs in reproducing historical climate variability, as evaluated by the Pearson correlation coefficient (r) and Nash–Sutcliffe efficiency (NSE). Among them, the LSTM method exhibited the highest overall accuracy and best capability in capturing temporal variability, and was thus selected to integrate future precipitation and temperature projections under the three Shared Socioeconomic Pathways (SSPs). Projections under the SSP1–2.6, SSP2–4.5, and SSP5–8.5 scenarios indicate that both precipitation and temperature will rise relative to the baseline period (1979–2014), with the strongest increases under SSP5–8.5. Winter precipitation exhibits the most pronounced seasonal increase, while seasonal differences in temperature rise are less distinct, with slightly stronger warming in winter. The machine learning–enhanced ensemble framework proposed in this study improves regional climate projections and provides a practical tool for guiding hydrological impact assessments and climate adaptation strategies in alpine basins.