Abstract: Reference evapotranspiration (ET0) is one of the most important parameters to calculate the crop water demand. The key physical quantity of the water cycle can also pose a great challenge to the water balance for the decision-making on the agricultural water use plan at present. The ET0 in the facilities can be generally estimated by the improved empirical formula with high accuracy. But, much more meteorological parameters are required during estimation. It is prohibitively expensive for the experimental cost of ET0 estimation using measurement instruments for the meteorological parameters in the shading facilities. Alternatively, machine learning can be expected to easily obtain the meteorological parameters outside the facilities. However, only a few studies were focused on the estimation of ET0 in this case. In this study, an efficient and accurate estimation of the ET0 was proposed to clarify the relationship between the meteorological parameters inside and outside of Panax notoginseng shading facility. A Sobol sensitivity analysis was implemented to determine the effective meteorological parameters outside the facility as the model input. A Penman-Monteith model was used to calculate the standard values. Three ET0 estimation models (BO-SVR, BO-RF, and BO-ELM) were established, where the Bayesian Optimization (BO) was used to optimize the parameters in the Support Vector Regression (SVR), Random Forest (RF), and Extreme Learning Machine (ELM). The results showed that there was a strong correlation between six meteorological parameters inside and outside the shading facility, among which the average temperature, the maximum temperature, the minimum temperature, and average relative humidity were significantly correlated, and the coefficient of determination (R2) values were 0.914, 0.721, 0.925 and 0.923, respectively. The radiation term was close to 0 in the shading facility. The ET0 was approximately equal to the aerodynamic term, where the R2, the Root-Mean-Square Error (RMSE), and the Mean Absolute Error (MAE) were 0.999, 0.008 mm/d, and 0.006 mm/d, respectively. There was a strong correlation between the aerodynamic terms inside and outside the shading facility, where the R2, RMSE, and MAE were 0.856, 0.097 mm/d, and 0.073 mm/d, respectively. Therefore, it was feasible to estimate the ET0 in the shading facility of Panax notoginseng using the meteorological factors outside the facility. In Sobol sensitivity analysis, the ET0 in the shading facility was highly sensitive to the average relative humidity, average wind speed, maximum temperature, and average temperature, with the first-order sensitivity coefficients of 0.450, 0.304, 0.064, and 0.026, respectively. There was a small influence of the minimum temperature and sunshine duration on the ET0, where the first-order sensitivity coefficients were less than 0.01. Therefore, an optimal combination of four meteorological parameters was constructed for the improved model. The overall performance of the BO-ELM model in the test accuracy was better than those of the BO-SVR and BO-RF models. The highest accuracy was achieved for the BO-ELM model using the optimal combination of average relative humidity, average wind speed, the maximum temperature, and average temperature, particularly with the R2, RMSE, and MAE of 0.928, 0.069 mm/d, and 0.046 mm/d, respectively. The BO-ELM model was also well adapted to estimate the ET0 in the facility with a small number of meteorological parameters (average relative humidity, and average wind speed), with the R2, RMSE, and MAE of 0.910, 0.078 mm/d, and 0.057 mm/d, respectively. The computational cost of each estimation model was calculated from the parameter tuning time and modeling time of the model. Overall, the BO-SVR and BO-ELM models presented relatively short running time of 0.97 and 2.22 s, respectively. By contrast, the longest running time of 27.46 s was obtained in the BO-RF model. Therefore, the BO-ELM model can be expected to serve as the ET0 estimation in the shading facility in the absence of some meteorological parameters, fully considering the calculation accuracy and cost of the simulation. The findings can also provide an effective way for the estimation of ET0 in the shading facilities.