基于主成分分析法优化广义回归神经网络的地震震级预测  

Earthquake Magnitude Prediction Based on Generalized Regression Neural Network Optimized by Principal Component Analysis

作　　者：王晨晖 袁颖 刘立申 陈凯南[1,2] 吴鹤帅 WANG Chen-hui;YUAN Ying;LIU Li-shen;CHEN Kai-nan;WU He-shuai(National Field Scientific Observation and Research Station for Huge Thick Sediments and Seismic Disasters in Hongshan,Hebei Province,Xingtai 054000,China;Xingtai Central Seismic Station,Hebei Earthquake Agency,Xingtai 054000,China;School of Urban Geology and Engineering,Hebei Geologic University,Shijiazhuang 050031,China;Hebei Province Underground Artificial Environment Intelligent Development and Control Technology Innovation Center,Shijiazhuang 050031,China)

机构地区：[1]河北红山巨厚沉积与地震灾害国家野外科学观测研究站,邢台054000 [2]河北省地震局邢台地震监测中心站,邢台054000 [3]河北地质大学城市地质与工程学院,石家庄050031 [4]河北省地下人工环境智慧开发与管控技术创新中心,石家庄050031

出　　处：《科学技术与工程》2022年第29期12733-12738,共6页Science Technology and Engineering

基　　金：国家自然科学基金(41807231);河北地质大学科技创新团队项目(KJCXTD-2021-08);河北省地震科技星火项目(DZ2021110500001)。

摘　　要：为科学有效预测地震震级,提出了基于广义回归神经网络(general regression neural network, GRNN)的地震震级预测模型。选取地震累计频度、累计释放能量、b值、异常地震群数、地震条带个数、活动周期、相关区震级等7个指标作为地震震级影响因子,利用主成分分析法(principal component analysis, PCA)对7个影响因子进行降维处理,以新生成的4个主成分作为模型输入变量,地震震级为输出变量,运用粒子群算法(particle swarm optimization, PSO)寻优得到GRNN模型最优光滑因子,最终建立基于PCA-PSO-GRNN的地震震级预测模型,利用建立的模型对训练样本进行回判检验,并对测试样本进行预测,并同传统反向传播(back propagation, BP)神经网络模型和单一GRNN模型预测结果进行对比,结果表明:PCA-PSO-GRNN模型预测结果的平均误差为5.17%,均方根误差为0.100 0,决定系数为0.986 8,均方相对误差为0.007 3,平均绝对误差为0.100 0,运行时间为5.2 s,预测精度和运行效率均优于BP模型和单一GRNN模型。In order to predict earthquake magnitude scientifically and effectively,earthquake magnitude prediction model based on general regression neural network(GRNN)was proposed.Seven indexes such as earthquake cumulative frequency,cumulative released energy,b value,number of abnormal earthquake clusters,number of seismic bands,activity cycle and magnitude of relevant areas were selected as the influence factors of earthquake magnitude.Principal component analysis(PCA)was used to reduce the dimension of the seven influence factors,and the newly generated four principal components were used as the model input variables,the earthquake magnitude was taken as the output variable,the particle swarm optimization(PSO)algorithm was used to find the optimal smoothing factor of GRNN model,and finally the earthquake magnitude prediction model based on PCA-PSO-GRNN was established.The PCA-PSO-GRNN model was used to test the study samples and predict the test samples,the prediction results were compared with results of the back propagation(BP)neural network model and the single GRNN model.The results show that the average error of the PCA-PSO-GRNN model is 5.17%,the root mean square error is 0.1000,the coefficient of determination is 0.9868,the mean square error is 0.0073,the mean absolute error is 0.1000,the running time is 5.2 s,and the prediction accuracy and operation efficiency were better than BP model and single GRNN model.

关 键 词：主成分分析法 粒子群算法 广义回归神经网络 地震震级 

分 类 号：P315[天文地球—地震学]