Evapotranspiration (ET) is a key component of water cycles and energy balance, quantification of which is important for revealing the eco-hydrological processes in arid and semi-arid areas. Taking the desert riparian Populus euphratica Olive. forest as the study object, we studied the diurnal and phonological variations of ET based on the data measured from the eddy covariance instrument, simulated ET of the P. euphratica forest using an improved dual-source Penman-Monteith-Priestley-Taylor (PM-PT) model and analyzed the parameter sensitivity. Results showed that: (1) the diurnal variations of ET of the P. euphratica forest both increased firstly, and then decreased during the growing seasons in 2014 and 2015, respectively. With the enhancement of solar radiation in the forenoon, the temperature raised gradually, ET increased and peaked at about 12:00. Afterwards, solar radiation weakened, the temperature decreased gradually, the relative humidity increased and the vapor pressure deficit decreased, and ET decreased accordingly. (2) The seasonal variations of ET of the P. euphratica forest increased firstly, and then decreased. Total ET during the growing seasons in 2014 and 2015 were 612 mm and 658 mm, respectively. ET accumulated in fruit and seed dispersal periods were the main parts, especially in fruit period. The accumulated ET were 316 mm and 348 mm, accounting for 51.65% and 52.87% of total ET in each growing season in 2014 and 2015, respectively. The average values of ET in seed dispersal period were slightly lower than those in fruit stage. The accumulated ET were 261 mm and 271 mm, accounting for 42.71% and 41.12% of total ET in each growing season in 2014 and 2015, respectively. Because the leaves of P. euphratica were not fully formed in leaf development stage and the leaf activity decreased gradually in leaf discoloration stage, the average values of ET in both stages were the lowest. (3) Compared with the traditional dual-source Shuttleworth-Wallace (SW) model, the improved dual-source PM-PT model is optimized in terms of the model structure and parameter numbers, and its accuracy is also higher. (4) The improved PM-PT model had the greatest sensitivity to net radiation.