SOM-RBF神经网络模型在地下水位预测中的应用  

Application of Combining SOM and RBF Neural Network Model for Groundwater Levels Prediction

作　　者：刘博[1,2] 肖长来[1,2] 梁秀娟[1,2] 

机构地区：[1]吉林大学环境与资源学院,长春130021 [2]吉林大学地下水资源与环境教育部重点实验室,长春130021

出　　处：《吉林大学学报（地球科学版）》2015年第1期225-231,共7页Journal of Jilin University：Earth Science Edition

基　　金：吉林省科技引导项目(20080543);高等学校博士学科点专项科研基金项目(200801830044);教育部国家潜在油气资源项目(OSR-01-07)

摘　　要：利用自组织映射(SOM)聚类模型优化径向基函数神经网络(RBFN)隐层节点的方法,减小了RBFN由于自身结构问题在地下水水位预测中产生的误差。采用SOM对已有样本进行聚类,利用聚类后的二维分布图确定隐层节点的数目,并根据聚类结果计算径向基函数的宽度,确定径向基函数的中心,由此建立SOM-RBFN模型。以吉林市丰满区二道乡为例,采用2000—2009年观测的地下水位动态资料,利用SOM-RBFN模型对地下水位进行预测,验证其准确性,并分别以5、7、10a的地下水位动态数据为研究样本建立模型,考查样本数量对预测结果的影响。研究结果表明:SOM-RBFN模型预测地下水水位过程中,均方根误差(RMSE)的均值为0.43,有效系数(CE)的均值为0.52,均达到较高标准,因此SOM-RBFN模型可以作为有效而准确的地下水水位预测方法;同时RBF7的RMSE和CE均值分别为0.38和0.68,结果优于RBF5和RBF10,这就意味着在模型计算中样本数量不会直接影响预测结果的精度。As the hidden units of radial basis function network（RBF）were optimized by the theory of self-organizing map（SOM）,the groundwater levels forecasting error range,due to its structural problems,could be reduced.With the two-dimensional feature map and clustering results of SOM,the number of hidden units,the position and the width of the radial basis centers can be easily determined.The SOM-RBFN model can be established.The accuracy of the model was verified by predictinggroundwater level at Erdao Town in Fengman District of Jilin City based on observed groundwater level from 2000 to 2009.In addition,dynamic data of groundwater level for five years（2005-2009）,seven years（2003-2009）,ten years（2000-2009）,are used as study samples and make forecast one by one,which can examine that if the sample size could influence the forecast result.The results prove that SOM-RBFN model can be used in groundwater levels dynamic forecasting,because the averages of RMSEand CE are 0.43 and 0.52,respectively,which are the relatively good outcomes.And,the averages of RMSEand CEof RBF7 are 0.38 and 0.68,whose results are better than RBF5 and RBF10.Therefore,it can be known that the amount of data cannot directly influence the accuracy of results.

关 键 词：地下水位预测 SOM RBF 神经网络 

分 类 号：P641[天文地球—地质矿产勘探]