Abstract: Earthquake prediction is a difficult problem in contemporary science, and applications of machine learning methods in the prediction have drawn intensive attention. Large earthquakes can cause huge casualties and economic losses, and are the main goals of earthquake prediction. We studies intermediate-term (one-year) earthquake prediction in Sichuan and Yunnan provinces using the earthquake catalogue since 1970 by the sliding time-space window technique and LSTM(long short-term memory) neural networks. Sixteen earthquake prediction indexes that reflect the temporal and spatial features of earthquake sequences were used in the neural network. The neural network was trained using data sets from 1970 to 2004 (70% of all earthquake catalogues). Retrospective prediction tests were conducted on earthquakes after 2005, the accuracy rate (actual magnitude fell within ±0.5 of the predicted magnitude) was 70.2%, over-prediction rate was 18.7%, and under-prediction rate was 11.1%. The 2008 Wenchuan M_S8.0 earthquake was retrospectively predicted. In order to understand the robustness of the model, we have done some tests, such as to expand the study area, change the weights of large earthquakes in calculation of the mean square error, etc. The LSTM neural network model still performed well in the tests.

Key words: intermediate-term earthquake forecast, long short-term memory neural network, earthquake prediction index, R value, Sichuan-Yunnan region