Abstract: A support vector regression machines model for real-time forecasting of reference evapotranspiration was developed, based on the ideas of structural risk minimization of support vector machines, as well as the global optimum capacity via the nonlinear mapping to the high-dimensional space of the kernel function. The model input included maximum, minimum and average daily temperatures, weather index and wind scale. The observation data during the period from 2003 to 2005 at Nanjing, Jiangsu Province were used as training and validating data set. A support vector regression model and an add noised support vector regression model by noise were presented, and were compared with a BP neural network model. The results showed that the proposed model had high accuracy (effectiveness index is 87.93%, the average error is 0.2609, the passing rate is 87.4%), compared with the BP neural model (effectiveness index was 78.91%, the average error was 0.2609, the passing rate was 76.8%), it has superior generalization capacity, the generalization ability will not reduce by noise, it has a strong adaptability and portability parameters.