基于投影寻踪的径向基函数网络在参考腾发量预测中的应用

Using radial basis function network based on project pursuit to forecast reference evapotranspiration

径向基函数网络(RBFN)已广泛应用于参考腾发量预测等领域,但常用的K-均值聚类和自组织特征映射等方法在求取径向基函数网络隐层节点中心时存在较大不足。针对这一问题,本文引入投影寻踪方法,在投影降维的基础上实现对大量高维数据的聚类,建立了基于投影寻踪的径向基函数网络模型,并将该模型应用于山西潇河灌区参考腾发量的预测,研究了不同气象因子输入对参考腾发量预测精度的影响。结果表明,基于投影寻踪的径向基函数网络具有较强的适用性,只需使用最高温度、最低温度、日照时数和旬序数作为输入因子,就能以较高的精度预测参考腾发量。

The K-mean value clustering and self-organizing feature map clustering methods in the application of radial basis function network （RBFN） are low efficient in searching the centers of RBFN hidden layer in forecasting of evapotranspiration when the training data are massive. The model of RBFN based on project pursuit is established and its learning process is described. The proposed model is applied to predict the reference evapotranspiration of Xiaohe Irrigation Experiment Station, Shanxi Province. The impacts of different combinations of input factors on the predicted reference evapotranspi- vation of RBFN are discussed. The results show that the new method is feasible. Using only maximum temperature, minimum temperature, sunshine hours and 10-day ordinal number as inputs, the model can forecast the reference evapotranspiration with a promising precision.