Abstract: This paper explores earthquake prediction using neural networks, focusing mainly on using long-short-time memory (LSTM) neural networks to construct an earthquake prediction model. Based on the Southern California earthquake catalog data from 1932 to 2021, the earthquake catalog from January 1932 to March 2002 was used as the training set(80% of the entire earthquake catalog), and the earthquake catalog from March 2002 to September 2021 was used as the test set (the remaining 20%). The LSTM neural network was selected, and 11 earthquake prediction factors reflecting the spatiotemporal intensity distribution characteristics of the earthquake time series data were calculated from the training set. The maximum magnitude label corresponding to these factors was used to construct the model. The test set was then used for retrospective prediction testing. The model’s prediction performance was evaluated using metrics such as accuracy, precision, and R-value, which were calculated based on the values in the confusion matrix. The results show that the prediction has achieved certain results, predicting the M7.2 earthquake in April 2010. The R-value of some models is significantly higher than China’s current medium-term prediction level. However, the value of the evaluation model is still not ideal, and further exploration is needed.

Key words: LSTM neural network, Southern California, medium-term earthquake prediction, seismic factor, R-value