[S-SS03] New insights in Earthquake predictability: modelling and forecasting

[SSS03-P03] Machine Learning Models for Earthquake Number Forecasting in the Kanto Region

*Hanyuan Huang¹, Hiroe Miyake¹, Hiroshi Tsuruoka¹ (1.Earthquake Research Institute, University of Tokyo)

Keywords:earthquake number forecasting, machine learning, long short-term memory

Machine learning techniques are developing in seismology contributed by the increasing amount of observable information. Some researchers suggest the earthquake cannot be predicted while many others think it is a predictable phenomenon. Through years of research, some regularities and patterns are found in the spatial and temporal distribution of seismicity. This research focuses to forecast the earthquake number on a certain day by using machine learning techniques to analyze past earthquake catalog, which could be considered as a reference for future earthquake hazard evaluation. We obtained the earthquake catalog of the Kanto region (longitude from 34.5 to 37.0, latitude from 138.5 to 141.5) from the Japanese Meteorological Agency (JMA), from 1 January 1999 to 31 December 2019. And the earthquake catalog is converted to statistical data in the patterns of the summation, mean, median, and max value of earthquake magnitude and number in different time windows (1 day, 3 days, 5 days, 7 days, 1 month, 2 months, 3 months, and one year) before the targeted day. As the machine learning models, multiple linear regression (MLR), and a 9 layers recurrent neural network (RNN) containing full connected layers, convolution layers, and long short-term memory (LSTM) layers are applied. Considering that the different patterns may exist in different depth and magnitude, we built different datasets containing earthquakes whose JMA magnitude (M_J) are greater or equal to 2.0 (M_J 2.0 sets), 3.0 (M_J 3.0 sets), and 4.0 (M_J 4.0 sets), with depths in the range of 0–30 km, 0–100 km, and 30–100 km respectively. The ratio of training, validation, and test sets is 6:2:2, and each model is trained separately. In the sector of results, because the recorded catalog with magnitudes of 2.0 or larger in 0–100 km set is abundant almost linearly related to the time except at around December 2016 when a shallow M_J 6.3 earthquake occurred, both MLR and RNN models perform pretty well, but in 0–30 km set, the MLR model almost totally fail to fit the trending while the RNN model still yields good forecasting. Among all the test tests, the results yielded from the MLR model are unstable while the performance of the RNN model is always satisfactory. Compared the results from machine learning models with the epidemic type aftershock sequence (ETAS) model, in all tested sets (M_J 3.0 sets and M_J 4.0 sets), the RNN model yielded the best results in cumulative number forecasting ranging from 1 January 2016 to 31 December 2019 while each the ETAS model and the MLR model reaches the second tier at times. For instance, the observed number of earthquakes with magnitudes of M_J 3.0 or larger with depth 0–100 km is 1519, and the forecasted value from RNN, MLR, and ETAS model is 1408, 1166, 1967. Yet the RNN model’s performance right after a relatively large earthquake is worse than others, and the MLR model lacks stability prominently. To conclude, we applied a recurrent neural network to forecast the number of earthquakes in the Kanto region. Even though the large magnitude earthquake number forecasting in a short term time-window (one day) seems impossible, the RNN model can output more accurate results in a long term time-windows cumulative number forecasting compared with the ETAS model, and this result may indicate that the earthquake temporal distribution can be better described by deep layer machine learning models.