基于支持向量机的现地地震预警地震动峰值预测

以0.1~10Hz带通滤波后三分向矢量合成地震动峰值PGA与PGV为预测目标参数,利用日本K-net强震台网P波触发后3s数据,基于人工智能中的经典机器学习方法-支持向量机,选取加速度幅值Pa、速度幅值Pv、位移幅值Pd、傅里叶谱幅值AMmax、速度平方积分IV2、破坏烈度DI、累积绝对速度CAV、阿里亚斯烈度Ia这8种特征参数构建地震动峰值预测模型。结果表明,对比常规的Pd预测模型,建立的支持向量机PGA与PGV预测模型,在测试数据集及随机选取2次震例数据集上计算得到的预测值与实测值更趋近1∶1比例关系,且PGA与PGV的预测值误差不受震中距变化的影响,PGA与PGV预测时的低值高估与高值低估现象也得到了改善。构建的支持向量机预测模型可用于现地地震预警地震动峰值、即仪器地震烈度的预测。

Prediction of peak ground motion for on-site earthquake early warning based on SVM

Here,taking ground motion peak values PGA and PGV synthesized by 3-D vectors after 0.1-10 Hz band-pass filtering as the prediction target parameters,using the data of 3 seconds after triggering P wave of Japan K-net strong earthquake network,based on the classic machine learning method in artificial intelligence-support vector machine(SVM),a ground motion peak value prediction model was constructed by selecting 8 feature parameters including acceleration amplitude Pa,velocity amplitude Pv,displacement amplitude Pd,Fourier spectrum amplitude AM max,velocity square integral IV2,destruction intensity DI,cumulative absolute velocity CAV and Arias intensity I a.Results showed that compared with the conventional Pd prediction model,using the constructed SVM prediction model for PGA and PGV,the predicted values calculated with testing data set and data sets of 2 earthquake cases selected randomly and the actual measured values are closer to a 1∶1 relationship,the errors for the predicted values of PGA and PGV are not affected by the change in epicenter distance;phenomena of under-estimation and over-estimation of PGA and PGV during prediction are improved;the constructed SVM prediction model can be used to predict ground motion peak values for on-site earthquake early warning,i.e.,to predict the instrument seismic intensity.