Evapotranspiration (ET) plays an important in the exchange of water, carbon and energy on land. The ET estimates based on different models and remote sensing data have different degrees of uncertainty. Bayesian model averaging (BMA) provides a way to reduce the uncertainty. Based on water and heat flux observation data in the Three-River Headwaters Region of China, evapotranspiration simulated by ARTS and PT-JPL models, and internationally shared MOD16 and SSEBop remote sensing evapotranspiration products, we conducted a BMA integration study, resulting in a dataset of surface evapotranspiration in the Three-River Headwaters Region of Qinghai Province based on Bayesian model averaging from 2003 to 2015. By verifying the results of each input model and BMA integrated model, it is found that the correlation between ET based on BMA and flux observation data is 0.94, which can explain 89% of the seasonal variation data, outperforming the performance of a single model. The results show that BMA model integration can integrate the inherent advantages of different models and reduce the uncertainty of result estimation for more reliable estimation results. This dataset can provide more accurate data for studying hydrothermal changes and evaluating ecosystem regulation functions in the Three-River Headwaters Region.