GRNN神经网络在矿区地表变形预测中的应用  

Application of GRNN in the Prediction of Surface Deformation in Mining Areas

作　　者：高彩云[1,2] 崔希民[1] 高宁[2] 洪雪倩[1] 

机构地区：[1]中国矿业大学地球科学与测绘工程学院,北京100083 [2]河南城建学院测绘工程学院,河南平顶山467036

出　　处：《金属矿山》2014年第3期97-100,共4页Metal Mine

基　　金：国家自然科学基金项目(编号:41071328)

摘　　要：针对矿区地表变形预测受多种因素影响的复杂性、非线性等特点,基于新型广义回归神经网络(GRNN),构建了矿区地表变形预测模型。首先,介绍了GRNN的建模原理,并对影响GRNN网络预测的关键因素进行了讨论;其次,为了提高网络的泛化能力及预测精度,采用滚动建模方式对网络进行建模训练,并基于最小均方误差原理提出了交叉验证搜索算法对GRNN网络预测关键参数平滑因子SPREAD进行优选;最后,将优化后的GRNN网络应用于某矿区地表变形预测,并与LM-BP、RBF、回归分析3种模型的预测效果进行了比较,结果表明,GRNN网络泛化能力强、算法稳定,且预测精度较高,适合于矿区地表变形预测。In view of the complexity and nonlinear characteristics of the prediction results, a new prediction model of sur- face deformation in mining areas is constructed based on the generalized regression neural network （GRNN）. Firstly, the model- ing principles of GRNN are discussed and the key factors that affect the prediction accuracy of GRNN model are introduced. Then, in order to improve the generalization ability and prediction accuracy of the network, the network is modeled and trained by adopting the rolling modeling method. The optimal smoothing factor SPREAD is determined in line with the across validation algorithm based on RMSE. Finally, the optimized GRNN is applied to predict the surface deformation in a mining area. The prediction results of BP neural 1 network based on Levenberg-Maquardt algorithm, RBF neural network and re- gression analysis method are used to compare with one of the optimized GRNN. The results show that, the GRNN net work gen- eralization ability and prediction accuracy are better than the others, in addition, the algorithm of optimized GRNN is stable. So, the optimized GRNN is suitable for surface deformation prediction in mining areas.

关 键 词：矿区地表变形 GRNN神经网络 滚动建模 交叉验证 预测 

分 类 号：P207[天文地球—测绘科学与技术]