A traditional multiple linear regression model （MLR） and an artificial neural network model （Back propagation （BP） and Elman） were constructed using 8 combinations of 7 meteorological factors including sunshine duration， wind speed， relative humidity， saturated vapor pressure difference， minimum temperature， maximum temperature， and average daily temperature as model inputs to solve the problems of low accuracy in estimating the daily transpiration （T_r） of Populus alba var. pyramidalis Bunge， the main tree species of shelterbelt forests in arid areas， and the poor generalization ability of estimation models. The T_r value of Populus alba var. pyramidalis Bunge in 2020 growth season was estimated. The results of estimating three different input combinations of models were compared and evaluated. At the same time， the relative contribution rates of various meteorological factors to the estimated T_r values were quantified based on the Garson algorithm. The results showed that the accuracy of BP and Elman models in estimating T_r exceeded 73.66%. Under different input combinations， the estimation accuracy of the artificial neural network model had increased by 8.45%-31.33% compared to the MLR model. Among them， the Elman model with a topological structure of 6-4-4-1 had the highest accuracy of estimation. The relative contribution rate of saturated vapor pressure difference to T_r estimation was the largest， followed by relative humidity. The relative contribution rate of different temperature variables to the estimated T_r values was in the increasing order of average daily temperature > minimum temperature > maximum temperature. It is indicated that the neural network model for estimating daily transpiration of Populus alba var. pyramidalis Bunge can improve the accuracy of estimating the transpiration of shelter forests in arid areas. It will provide scientific guidance for the sustainable development of shelter forests and the precise regulation of agricultural water resources.