In this study, by selecting the wetland, farmland, grassland, tamarix chinensis, populus euphratica forest and mixed forest ecosystem in the Heihe River Basin as studying objects, several machine learning algorithms were chosen to simulate and interpolate the latent heat flux data, considering several meteorological factors including the net radiation, temperature, soil heat flux, wind speed, relative humidity, and volumetric water content of soil. These methods included multiple linear regression(MLR), decision tree(CART), random forests(RF), support vector regression(SVR), BP artificial neural network(BPANN), and deep learning(DL). The results show the following: (a)RF, SVR, BPANN and DL obtain the best results in different ecosystems with R² = 0. 8 to 0. 93, R_MSE=21. 730 to 41. 731 W/m², and M_AE=12. 153 to 26. 129 W/m², but the results of SVR in tamarix and mixed forest ecosystems are slightly worse than other three methods with R² decreased by 0. 01 to 0. 02. The results of MLR are the worst with R² =0. 6 to 0. 7 and the results of CART are in between with R²=0. 78 to 0. 9. (b)By comparing whether soil moisture participating in the gap filling, it indicates that the participation of soil moisture can improve the gap-filling accuracy of models to some extent with R² increased by 0. 01 to 0. 06. (c)At the same time, the established gap-filling model was used to interpolate the evapotranspiration of other years, and it is found that the accuracy of gap-filling result decreased in varying degrees. With comprehensive consideration of the accuracy and stability of these models, it can be found that RF, BPANN and DL are more suitable for the gap filling of evapotranspiration, and the participation of soil moisture can improve the accuracy of three models.