基于深度神经网络的地表地震动幅值预测研究

Study on the Prediction of Ground Motion Amplitude Based on Deep Neural Network

地震动预测模型是灾害分析和结构设计的重要组成部分,近年来神经网络技术愈发成熟地被应用在该预测模型的开发上,但多为使用地表以下30 m范围内土层等效剪切波速(VS30)作为场地输入参数的单层全连接神经网络模型,忽略了完整土层厚度及剪切波速信息对地震动幅值的影响。本文采用卷积神经网络及全连接神经网络混合模型,选用日本KiK-net台网记录到的3174次地震共计39192条地震动记录,构建了一种基于深度神经网络框架的地表地震动幅值预测模型。该模型的输入参数为震级、震源深度、震中距、场地各土层厚度、剪切波速信息和井下地震动幅值,输出为相应的地表地震动幅值(PGA或PGV或PGD)。对模型进行训练并计算其各项评价指标予以评估,结果表明:(1)该混合神经网络模型的决定系数超过了0.85,模型残差服从正态分布,均值残差接近于0,模型表现出无偏的特性。(2)与现有经验公式相比,混合网络模型的PGA、PGV和PGD预测精度分别提升约26.9%、16.5%和11.6%。与使用VS30作为场地参数的全连接神经网络模型相比,该框架下模型预测值与真实值的Person相关系数及各项评价指标均有所提升,模型残差的均值和标准差更小,PGA、PGV和PGD模型预测精度提升约6.3%、3.9%和3.4%,能更好地对地震动幅值进行预测。

The ground motion prediction model is a critical component in disaster analysis and structural design.In recent years,neural network technology has increasingly been applied to develop such prediction models.However,most existing models are single-layer fully connected neural networks that use the equivalent shear wave velocity of soil layers within 30 meters below the surface(VS30)as the site input parameter.This approach often overlooks the impact of the complete soil layer thickness and shear wave velocity information on ground motion amplitude.In this paper,a combined convolutional neural network(CNN)and fully connected neural network model is proposed.A dataset comprising 39192 ground motion records from 3174 earthquakes recorded by KiK-net is used to build a ground motion amplitude prediction model based on deep neural network technology.The model's input parameters include the magnitude,source depth,epicenter distance,thickness of each soil layer,shear wave velocity information,and underground ground motion amplitude.The output is the corresponding ground motion amplitude(PGA,PGV,or PGD).The model is trained and its evaluation indexes are calculated.The results indicate that:(1)The coefficient of determination(R2)of the mixed input neural network model exceeds 0.85.The model residual follows a normal distribution,and the mean residual is close to 0,demonstrating an unbiased characteristic.(2)Compared to the empirical formula,the prediction accuracy of the hybrid network model improves by 26.9%,16.5%,and 11.6%for PGA,PGV,and PGD,respectively.(3)Compared to the traditional fully connected neural network model using VS30 as the site parameter,the Pearson correlation coefficient and various evaluation indexes of the predicted values and actual values show im-provement.The mean and standard deviation of the model residual are smaller,and the prediction accuracy of the PGA,PGV,and PGD models improves by about 6.3%,3.9%,and 3.4%,respectively.This combined neural network model can better predict seismic amplitude,taking into account the complete soil layer information.