Prediction Of Earthquake In Yunnan Region Based On The Support Vector Machine

As a new technology of machine learning based on statistical learning theory, support vector machine (SVM) apply structural risk minimization rule and kernel function theory, and it can serve better in the processing of small-sample, high dimensions and non linearity problems. Therefore, SVM theory has been obtained intensive study in recent years and has been employed in various fields such as monitoring, predicting, modeling and controlling etc.SVM and its some other improved algorithms are used in this paper to predict the date and magnitude of earthquakes in Yunnan region, and the prediction result is verified. The main research content is shown as follows:(1) By collecting and compiling the date and magnitude of Yunnan region over the years, we extract the seismicity indicators which reflect the information of earthquake activity, and construct the raw data set. Then the samples are divided into positive and negative class according to the threshold we set, and classification method is used for prediction. When the threshold is small, we can predict the small earthquake or the medium one. Against the imbalance of the positive and negative samples, we present the method of combination of AHC over-sampling and biased SVM (AHC-BSVM) to classify and predict. The experiment result shows the AHC-BSVM can obtain a higher classification accuracy and better prediction result than the conventional and standard SVM and biased SVM. And it also shows that it has higher accuracy and lower false alarm rate to predict the small earthquake than to predict the medium one by AHC-BSVM.(2) When the threshold is large, we can predict the strong earthquake. Aimed at the extremely imbalance of the positive and negative samples, one-class classification SVM is employed to classify and predict. In the first place, the main theory of conventional spherical one-class classification SVM (SOCC) and the advanced hyperellipsoid one-class classification SVM (HOCC) are elaborated. Then the strong earthquake is predicted by AHC-BSVM, SOCC and HOCC respectively. Under the extremely imbalance of the positive and negative samples, experiment shows that HOCC obtains better prediction result than SOCC and AHC-BSVM. Applying HOCC for predicting strong earthquake can improve the accuracy and lower the false alarm rate. Besides, RBF kernel function performs better than the polynomial kernel function in using the same classifier.(3) When the regression method is used for prediction, we consider a year as a waited-prediction period. Then the maximum magnitude over the years in Yunnan region is collected as a time series andε-support vector regression (ε-SVR) is introduced for prediction, then the prediction result is verified. Experiment shows the input dimension of the prediction model and the interrelated parameters of theε-SVR have a great influence to the prediction result, and we need to employ some specific evaluation metrics and optimization rule for obtaining a better performance of the prediction model. Besides, in the case of setting appropriate parameter,ε-SVR performs better than the BP neural network. Though the waited-prediction period ofε-SVR is longer, it obtains higher accuracy than the classification method of AHC-BSVM or HOCC.Though the earthquake prediction model proposed in this paper couldn't predict all of the earthquakes and appear some false alarms and failed alarms, it can identify and predict certain earthquakes in a specified magnitude range. So it provides some references for disaster prevention and reduction, and explores a new way for earthquake prediction with machine learning.