As the only connecting parameter between energy balance and water balance, evapotranspiration (ET) is the most excellent indicator for the activity of climate change and water cycle, and therefore, accurate estimation of ET is of importance for hydrologic, climatic and agricultural studies. ET is commonly computed by reference evapotranspiration (ET_0), and this paper investigated the performance of generalized regression neural network(GRNN) algorithm in modeling FAO-56 Penman-Monteith(PM) ET_0 only with the temperature data in 19 meteorological stations located at the western, middle, and eastern part of Sichuan basin, southwest China. Data of meteorological variables containing maximum air temperature (T_(max)), minimum air temperature(T_(min)) for the period of 1961-1990 were used as input variables to train the GRNN model, and data for the period of 1991-2014 were used to validate the GRNN model. The performance of GRNN model was compared with the empirical temperature-based Hargreaves(HS1) and calibrated Hargreaves(HS2) models considering the PM ET_0 as the benchmarks. The evaluation criteria of root mean squared error(RMSE) and model efficiency (E_(ns)) were used for the comparison. The statistical results indicated that the RMSE values of the GRNN, HS1 and HS2 models were 0.41, 1.16 and 0.70 mm/d, respectively, and the E_(ns) values of the GRNN, HS1 and HS2 models were 0.88, 0.13 and 0.67, respectively, which manifested the performance of the GRNN model was encouraging. The RMSE of the HS1 model was the biggest every year at temporal scale in all 3 sub-zones of Sichuan basin, followed by the HS2 model, and the RMSE of the GRNN model was the smallest; the E_(ns) of the GRNN model was bigger than HS1 and HS2 model every year at temporal scale in all 3 sub-zones. The RMSE of the HS1 model was the biggest in every meteorological station of Sichuan basin at spatial scale, followed by the HS2 model, and the RMSE of the GRNN model was the smallest; the E_(ns) of the GRNN model was bigger than HS1 and HS2 model in every meteorological station at spatial scale. Based on the RMSE and E_(ns), the errors of GRNN, HS1 and HS2 models showed an increasing tendency, which indicated the error of all the 3 models would become bigger in the future. The ranges of the ET_0 values computed by the PM, GRNN, HS1 and HS2 model were 695~837, 709~820, 1 029~1 178 and 818~975 mm, all of which showed an increase tendency with a rate of 2.7, 2.0, 2.2 and 2.4 mm/a respectively in recent 24 years. Compared with the PM model, GRNN, HS1 and HS2 overestimated the ET_0 value by 0.8%, 45.1% and 17.3%, respectively. The analysis of the performance of GRNN, HS1 and HS2 models at temporal and spatial scale confirmed the good ability of the GRNN model in estimating ET_0 when the data for PM model were not fully available, and thus the GRNN model should be adopted to compute ET_0. This paper can provide the reference for estimating the crop water requirement in Sichuan basin when only temperature data are accessible.