Assessing climate change impacts on streamflow in upper Han River Basin using deep learning models ensembled with Bayesian model averaging
Xin Wang, Chao Deng, Xin Yin, Jia Wei, Jia-cheng Zou
Abstract
Accurate streamflow prediction under climate change is essential for mitigating natural disasters and optimizing water resources management. However, streamflow prediction is subject to considerable uncertainties due to the complexity of hydrological model structures, parameterization, and input forcing data. This study predicted monthly streamflow in the upper Han River Basin in China under three Shared Socioeconomic Pathways (SSP) scenarios, using climate projections from five Coupled Model Intercomparison Project Phase 6 (CMIP6) climate models. Bias correction of climate model outputs was performed prior to streamflow simulation using four deep learning approaches: long short-term memory, gated recurrent unit, temporal convolutional network, and transformer. To reduce uncertainties inherent in individual deep learning models, Bayesian model averaging (BMA) was employed to integrate their predictions. The results showed that the three deep learning models achieved satisfactory performance with Nash–Sutcliffe model efficiency coefficient (NSE) values exceeding 0.8, while BMA exhibited superior robustness and accuracy, with the highest NSE and lowest root mean square error. Projected precipitation, mean air temperature, and potential evapotranspiration generally decreased during 2026–2100 relative to the historical period (1970–2017), suggesting a colder and drier regional climate. Streamflow was projected to decline significantly across all three scenarios, particularly from June to September, highlighting the potential for exacerbated water scarcity in the future.