To effectively improve the prediction accuracy of the reference crop evapotranspiration (ET0) in the arid regions of Northwest China, five representative meteorological sites were selected in the arid Northwest China to construct 10 errors back propagation neural network (BPNN) optimized by mind evolutionary algorithm (MEA) model. This model was used to forecast ET0 and compared with the three models of Hargreaves-Samani model, Irmak model and 48-PM model which had higher accuracy in the northwest arid region. The results showed that the simulation accuracy of the MEA-BPNN model was basically high at different input levels, including MEA-BPNN1 (input Tmax, Tmin, RH, n and u2), MEA-BPNN2 (input Tmax, Tmin, n and u2) and MEA-BPNN3 (input Tmax, Tmin, RHand u2). The determination coefficient R2 and Nash-Sutcliffe efficiency coefficient NSE of the models were greater than 0.96, RMSE and MAE was less than 0.34mm/d and 0.25mm/d. The GPI rankings of the above three MEA-BPNN models were 1, 2 and 3, respectively. The R2 and NSE of MEA-BPNN7 (input Tmax, Tmin, and u2) was 0.966 2 and 0.962 2, RMSE and MAE was 0.361 0mm/d and 0.2761mm/d, respectively, and the simulation accuracy was high. The analysis of the portability of the MEA-BPNN model showed that the MEA-BPNN model in the arid northwestern China had strong generalization ability, and the forecasting model constructed based on different site data also had high accuracy. The simulation accuracy of the MEA-BPNN model was higher than that of the Hargreaves-Samani model, Irmak model and 48-PM model with the same input. Therefore, in the absence of meteorological data, the MEA-BPNN model can be used as a recommended model for the calculation of ET0 in the northwest arid regions, which can provide a scientific basis for realtime accurate irrigation forecasting.