Prediction Of On Reference Crop Evapotranspiration Based On Back-Propagation Network

In order to overcome the complexity in the calculation of reference evapotranspiration (ETo) in the agriculture soil and water engineering discipline, back-propagation (BP) neural network which has been proved to be powerful tools in modeling nonlinear process, was applied to the calculation and prediction of ETo. Based on the analysis of influence factors of ETo, BP neural network models with different structure have been established. Take the climate data and rice evapotranspiration of Yujiang Jiangxi province in 2003 as training sample, and the data in 2004 as verify sample. The models were trained and evaluated, and the characteristic and applicability of each model were illustrated. The major contents and results are listed as follows:1. Based on the analysis of influence factors of ETo and the characteristic and shortcoming of BP neural network, 3 models including INALL9-BP model (with all climate data available in weather stations as inputs), WFbased-BP model (with climate data available in weather forecast as inputs) and TScombined-BP model (with both data available in weather forecast and time series as inputs) were established according to different regimes of available climate data. Models structures were determined.2. Take the climate data and rice evapotranspiration of Yujiang Jiangxi province in 2003 as training sample, and the data in 2004 as verifying sample, the models were trained and evaluated. All models are acceptable both in the simulation and prediction of ETo and the rice evapotraspiration. It also demonstrated that the INALL9-BP model performed much better than others, and WFbased-BP model was less useful.3. According to the evaluation of models simulation and prediction of ETo and the rice evapotraspiration, INALL9-BP model was recommended firstly when all the climate data are available. But when the climate data are absent, TScombined-BP model which combined the time series and data available in weather forecast was recommended as first choice, which can give acceptable results with limited inputs.4. The characteristic and applicability of each model were discussed. Results show the 3 layers neural networks are more sensitive to the number of neural in hidden layer than 4 layers neural networks, but the prediction results of 3 layers neural networks may be better than the 41ayers neural networks when the numbers of neuralin hidden layer are within a reasonable section. It was suggest that the layers number of neural networks should be determined when there are different requirement on the sensitivity and error control of both the simulation and prediction results.