Abstract: The occurrence of an earthquake is instantaneous and destructive, which may cause serious disaster to human society. Therefore, the development of relevant theories and methods for earthquake prediction is of great significance. In this study, classical back propagation (BP) neural network and variational mode decomposition (VMD) technology are combined to develop a new VMD-BP earthquake prediction model. Four VMD-BP models were constructed to predict the time, magnitude, and location (latitude and longitude) of earthquakes. For each entry in the earthquake catalog, four features other than the depth of the earthquake source are considered: time difference, magnitude, latitude, and longitude. The past five adjacent seismic events was used as the input to the VMD-BP model, and the next seismic event was used as the output. The model was empirically analyzed in the Sichuan and Xizang regions. By comparing the prediction performance of the models, it is concluded that the VMD-BP model has better prediction accuracy and goodness of fit than the BP model. It illustrates that the VMD-based model can make more accurate earthquake prediction, which will provide useful references for future in-depth research on earthquake prediction.

Key words: Earthquake prediction, BP neural network, Variational mode decomposition, Deep learning