Research On Rarthquake Magnitude Prediction Based On Improved GWO Optimized SOM-SVR Algorithm

As one of the most destructive natural disasters,earthquakes have brought tragic lessons to human beings,and effective earthquake prediction can reduce casualties and economic losses.There are a wide variety of factors leading to earthquakes but there are nonlinear relationships between their factors.With the rapid development of computer technology,it makes machine learning models to solve many complex and nonlinear problems widely into applications.This paper uses MATLAB and python platform to improve the prediction accuracy by using self-organizing feature map neural network(SOM)and support vector regression(SVR)in the field of earthquake magnitude prediction.This paper first introduces the current state of machine learning models in the field of earthquake prediction.In view of the large variability and dispersion of earthquake data,the combination of SOM and SVR models is used to improve the accuracy and stability of magnitude prediction.At the same time,an improved method is proposed.The gray wolf algorithm(GWO)optimizes the SVR parameters,and the improved gray wolf algorithm solves the defect of easily falling into local optimum in the process of parameter optimization.In the selection of the study area,this article introduces in detail that Sichuan is a multi-seismic area and explains the reasons for the multi-seismic area.For extracting the input feature vector of the model,by consulting a large number of relevant documents,some precursor earthquake prediction factors are selected as the input of the model,and normalized processing is carried out to ensure the convergence of the model and the prediction accuracy.For the problem of large dispersion of seismic data samples,this paper adopts SOM for clustering to ensure the accuracy of the combined model.SVR model,as a tutored learning algorithm developed from the basis of statistical theory,is based on the basic principle of minimizing structural risk and improving the generalization ability of the model as much as possible.The SVR model shows strong advantages in solving both small and nonlinear samples,and for seismic data with few samples,the choice of SVR model has great advantages.Therefore,this paper proposes a new SOM-SVR prediction model combining SOM network without tutor learning and SVR with tutor learning,which not only meets the requirements of support vector machine for training sample size,but also ensures the accuracy of prediction.Although the SVR has good prediction accuracy,the SVR parameters and have a large impact on the prediction results,and it is crucial to get an optimal parameter for the output results.In this paper,we propose an improved Gray Wolf algorithm to perform global optimization of the SVR parameters C and g.The improved GWO algorithm is used to obtain an optimal parameter for the SVR model,which ensures the output of the combined model.Finally,the output results of the combined model(SOM-GWO-SVR)are processed by inverse normalization and also compared with the output results of the traditional SVR to obtain error plots of the output results of several models.On the one hand,it shows that the combined model proposed in this paper is feasible in earthquake prediction research;on the other hand,it verifies that the model proposed in this paper has higher accuracy and stronger robustness compared with other combined models.This paper also discusses the application of the model to Taiwan,China,where earthquakes are more frequent,by transfer learning,and verifies the generalization ability of the model.