基于GMDH的地震液化场地侧向变形预测模型

Prediction model for liquefaction-induced lateral spread displacement based on GMDH

基于GMDH神经网络,建立了地震液化场地侧向变形的预测模型,并将其结果与传统反向传播BP、遗传算法优化BP、径向基函数RBF等神经网络的预测结果进行比较.结果表明,缓坡和临空面场地中,所提模型在训练样本集的拟合相关系数分别为96.43%和93.82%,模型准确度较高.对于缓坡场地,倾斜率、液化土层厚度与侧向变形成正相关关系,震中距、平均细粒质量分数则与其成负相关关系;对于临空面场地,高度与距离长度之比、液化土层厚度与侧向变形成正相关关系,平均细粒质量分数、平均粒径与其成负相关关系.通过实际工程应用发现,所提模型的预测结果与经典的Youd简化模型结果吻合较好,由此证明了其可靠性,可在高烈度地震区工程建设中应用与推广.

Based on a group method for a data handing(GMDH)neural network,a prediction model for seismic liquefaction-induced lateral spread displacement was established,and the results of the model were compared with those of the traditional back propagation(BP),the genetic algorithm optimized BP,and the radial basis function(RBF)neural networks.The results show that the fitting correlation coefficients of the proposed model in the training set for the gentle slope site and the free surface site are 96.43%and 93.82%,respectively,indicating that the precision of the model is higher.For the gentle slope site,the slope rate and the thickness of the liquefied soil layer show a positive correlation with the lateral spread displacement,while the horizontal distance from the site to the seismic energy source and the average mass fraction of fines exhibit a negative correlation.For the free face site,the ratio of the height to the distance length and the thickness of the liquefied soil layer show a positive correlation with the lateral spread displacement,while the average mass fraction of fines and the average mean grain size exhibit a negative correlation.Through practical engineering application,it is found that the predicted results of the proposed model are in good agreement with those of the classical Youd’s simplified model,which proves the reliability of the proposed model.It can be applied and popularized in the engineering construction of high intensity seismic areas.