A Short-Term Earthquake Prediction Model Based On Neural Network And Its Applications

This research is sponsored by Jiangsu Science and Technology Support Programme (social development) Project "Neural Network Based Short-term Earthquake Prediction Methods Study" (No.BE2009663), the National High Technology Research and Development Program of China (863 Program) "Research on neural network technique for spatial data mining" (No.2007AA12Z228), etc.Earthquake is one of the natural phenomena which has made huge catastrophe to the human society. Therefore many domestic and foreign experts have been conducting earthquake prediction research and have many meaningful progresses, especially long-term prediction of earthquake. However, earthquake prediction is still one of the world’s scientific problems to be resolved. At present, earthquake prediction is still at the early stages and the overall level is still not high. Especially, short-term and imminent earthquake prediction level is far from the social needs.The activity of earthquake is high frequency, wide distribution and high strength. It will bring severe disasters in our country. The strong earthquake zone has a certain relationship with the activities in our country and the confining of the block activity. But the relationship is nonlinear. In recent years, neural network technology has been widely used in the field of earthquake prediction for its good nonlinear mapping function, which could reflect the non-linear relationship between various precursors and the earthquake magnitude and time in the future. In this paper, based on the BP neural network technology, earthquake prediction methods have been researched. A regional short-term earthquake prediction model has been established to give the quantitive prediction of the maximum magnitude in the research region within the next forcast period. The main contents and conclusions are as follows:(1) Decades of research work and achievements of the earthquake zone division have been studied and summarized. Based on the current characteristics of China’s earthquake zone distribution, this paper aims to propose a division method which make a statistical seismic activity zone as one zone. Finally, Chinese mainland has been divided into 40 prediction regions.(2) Activity of earthquake is affected by many parameters. In earthquake prediction, which parameters to chose have a subjective sidedness. This paper uses the factor analysis method to chose the predict parameters. Factor analysis methos can objectively determine the weights of each parameter, which could avoid the subjective sidedness. And multicollinearity correlation analysis between the various parameters is supplemented. Then a number of significant rights are determined and can reflect the characteristics of regional activity parameters. Finally, based on integration of multiple linear regression analysis model and neural network technology, a short-term earthquake prediction model is established.(3) In this paper, neural networks technology is the foundation of the earthquake prediction research. Firstly, the principle, insufficient and improvement of neural networks technology are introduced. Then the genetic and neural networks fusion algorithms are used to get the structure optimization and improve efficiency, which have achieved good results in earthquake prediction.(4) In order to verify the reliability and practicality of the integration astronomical earthquake prediction model, combining carved out of the "National 40 Regional distribution of earthquake prediction," We selecte 2#,7#,15# and 20# four areas from the North earthquake zone, the southwest earthquake zone and Taiwan earthquake zone to established the prediction model. The period of the prediction model is three-month. Results show that the accuracy of short-term earthquake prediction model with neural networks technology has great improvement. The accuracies of the multiple linear regression models in the four areas are 1.43Ms, 0.83Ms,1.08Ms and 0.84Ms respectivily. The accuracies of networks models in the four areas are 0.82Ms,0.65Ms,0.73Ms and 0.64Ms respectively. Compared to the multiple linear regression models, the accuracies of networks models improve 42.3%, 22.7%,31.9% and 24.6% respectively. Earthquake prediction results show that the networks models proposed in this paper have broad prospect of application, great scientific value and profound significance for the future global earthquake warning system.