Abstract: Abstract: Evapotranspiration (ET) is the sum of soil water evaporation and plant transpiration. Accurate prediction of ET can provide information and decision making for irrigation management of crops and efficient use of agriculture water. Attention-based LSTM model (AT-LSTM) is widely used in machine translation, and speech recognition. However, there is still a gap in the application of ET simulation. In this study, an evapotranspiration simulation was conducted in Yanchi county Ningxia, China, with the limited environmental data, thereby to verify the feasibility and effectiveness of an AT-LSTM model for the high accurate ET. Air temperature, net radiation, relative humidity, soil temperature, and soil water content were selected as the influential factor of ET. They were half hourly environmental factors for Yanchi county from January 1, 2012 to December 31, 2017. The ET was calculated from the latent heat flux (LE). The data from 2012 to 2016 served as the training set, and the data in 2017 served as the test set. Different combinations of environmental factors were used as the inputs to construct the AT-LSTM model, compared with the Extreme Learning Machine (ELM), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) model on daily, monthly, and seasonal scale. The Root Mean Square Error (RMSE), Mean Absolute Error (MAE), correlation coefficient (R), and Global Performance Indicators (GPI) were used to evaluate the performance of models. Compared with ELM, SVM, and LSTM models, the simulation accuracy of AT-LSTM model changed very little, when the input environmental factors changed, where all data were better. When all meteorological factors were input, the AT-LSTM model had the best effect, with the RMSE of 0.013 mm/30min, MAE of 0.006 mm/30min, and R of 0.905. Specifically, the LSTM model was inferior to AT-LSTM model, with the RMSE of 0.014 mm/30min, MAE of 0.007 mm/30min, and R of 0.889, respectively. The simulation accuracy of ELM and SVM was lower than that of LSTM. When the air temperature, net radiation, relative humidity, soil temperature were input, the GPI of ELM and LSTM models were all 2, while the GPI of AT-LSTM model was 6. When the net radiation and soil temperature were input, the GPI of AT-LSTM model was 2, with the RMSE of 0.013 mm/30min, MAE of 0.006 mm/30min, R of 0.900. The simulation effect that produced by the combination of environmental factors was related to the selected model. When only net radiation was input, the RMSE of AT-LSTM was 0.014 mm/30min, with MAE of 0.007 mm/30min and R of 0.892. This model outperformed the LSTM, ELM and SVM models with all environmental factors input. When the net radiation was added as the input for the four models, the simulation accuracy of model was improved. When the SWC was used as the input, the RMSE of AT-LSTM model was 0.016 mm/30min, with MAE of 0.008 mm/30min, R of 0.859, while, the simulation accuracy of SVM, ELM and LSTM was much lower than AT-LSTM. Compared with other models, the simulation results of AT-LSTM model were closer to the real value on the daily and seasonal scale. The AT-LSTM model with the high simulation accuracy and strong stability can be used to solve the simulation and prediction problem of evapotranspiration in Yanchi county, where the excellent simulation results were achieved on the hourly, daily, monthly, and seasonal scales. It infers that the deep learning was suitable for the simulation and prediction of evapotranspiration. The influence of meteorological factors on ET was greater than that of soil factors in Yanchi county. The simulation effect can be ranked in order: the model of AT-LSTM > LSTM > ELM > SVM. Specifically, the SVM model had the same simulation effect as the ELM model in the period of high temperature in May and September and at 10am to 4pm in summer. The net radiation played a leading role in ET among these meteorological factors, while, the SWC less. The AT-LSTM model with only input of net radiation can achieve high accuracy, and thereby to serve as a simulation model with the input of missing environment factor.