BP与GM(1,1)预测隧道涌水对比分析和实证  

Forecast Water Gush in Tunnel Based BP and GM(1,1):A Case Study

作　　者：安永林[1] 彭立敏[1] 

机构地区：[1]中南大学土木建筑学院,长沙410075

出　　处：《科技导报》2008年第13期71-74,共4页Science & Technology Review

基　　金：铁道部科技研究开发计划课题(2005K002-D-3)

摘　　要：利用某隧道的涌水监测数据,对比BP神经网络和GM(1,1)灰色数列预测模型两种方法预测结果的差异,以考察其适应性与误差。结果显示,在小样本信息量少的情况下,GM(1,1)预测精度优于BP,但训练样本的精度低于BP;BP的预测结果同隐含层神经元个数密切相关,并存在一最优值;在监测数据较少时,对BP网络进行初始化和预测,每次训练样本的误差都满足要求,但预测值的误差大幅波动。研究表明,监测数据较少时,采用GM(1,1)较合适。通过分析小样本下产生上述结果的原因,提出了在有足够监测数据下,GM(1,1)用于中长期监测的改进方法(GM(1,1)展开或用Verhulst模型)、BP神经网络的改进方法(滑移窗口处理)。BPnetwork model and GM（1,1） model were applied to forecast water gushing in a tunnel based on in-site monitoring data, and by comparing the results. The applicability of the two methods is analyzed. Results show that： （1） with a small number of data, predicted results of GM （1, 1） are better than those of BP, but with respect to the relative error of the sample, BP is better, （2） the forecasting precision of BP has a close relation to the nerve cell number in the hiding layer and there is an optimal nerve cell number;（3） in case of a small number of data, after initializing and training the same BP network, the relative error of the sample is very small, but the forecasting results fluctuate dramatically. Therefore, GM（1, 1） is better in forecasting water gushing if the number of monitoring data is small. The reasons are analyzed and some improvements are proposed for BP network and GM（1, 1）.

关 键 词：隧道工程 涌水 BP神经网络 GM(1 1)灰色数列预测模型 

分 类 号：U457.5[建筑科学—桥梁与隧道工程]