Abstract Analysing earthquake data is essential for reliable forecasting and the development of effective early warning systems. The present study developed a preprocessed, declustered, and homogeneous global earthquake dataset spanning 1900–2024 using data from the International Seismological Centre (ISC) and the United States Geological Survey (USGS). Annual seismic energy time series were developed to represent the temporal evolution of global seismic activity. To handle the non-linearity and non-stationarity of energy time series, the ensemble empirical mode decomposition technique was applied, yielding intrinsic mode functions (IMFs). Among six IMFs, IMF1, IMF2, and IMF3 were selected through random forest-based feature importance and then served as inputs for machine learning models. Four models, namely support vector regression (SVR), multilayer perceptron (MLP), long short-term memory (LSTM), and temporal fusion transformer (TFT), were evaluated using root mean squared error (RMSE), coefficient of determination (R2 score), and mean absolute error (MAE). SVR achieved the best generalization (RMSE = 0.293, R2 = 0.854, MAE = 0.231), followed closely by MLP (RMSE = 0.288, R2 = 0.859, MAE = 0.235). LSTM showed adequate learning ability but moderate testing performance, while TFT exhibited strong training accuracy but weak testing generalization. These results emphasize SVR’s robustness for seismic energy forecasting in data-limited conditions. Research highlights <OrderedList> <ListItem> <ItemNumber>1.</ItemNumber> <ItemContent> A global earthquake catalogue spanning over 500 years was used to construct an annual seismic energy release time series. </ItemContent> </ListItem> <ListItem> <ItemNumber>2.</ItemNumber> <ItemContent> Empirical Mode Decomposition (EEMD) was applied to extract intrinsic mode functions (IMFs) capturing multi-scale seismic trends. </ItemContent> </ListItem> <ListItem> <ItemNumber>3.</ItemNumber> <ItemContent> Machine learning models, including SVR, MLP, LSTM, and TFT, were developed for forecasting future energy release using IMF components. </ItemContent> </ListItem> <ListItem> <ItemNumber>4.</ItemNumber> <ItemContent> Results demonstrate the feasibility of supervised learning for anticipating future global seismic energy release trends. </ItemContent> </ListItem> </OrderedList>

Data-driven prediction of global annual seismic energy using machine learning models

期刊论文
发表日期：

2026年3月12日

查看全文

Bala Balaji Yarramsetty,
Kavitha Baladhandapani

摘要

Abstract

Analysing earthquake data is essential for reliable forecasting and the development of effective early warning systems. The present study developed a preprocessed, declustered, and homogeneous global earthquake dataset spanning 1900–2024 using data from the International Seismological Centre (ISC) and the United States Geological Survey (USGS). Annual seismic energy time series were developed to represent the temporal evolution of global seismic activity. To handle the non-linearity and non-stationarity of energy time series, the ensemble empirical mode decomposition technique was applied, yielding intrinsic mode functions (IMFs). Among six IMFs, IMF1, IMF2, and IMF3 were selected through random forest-based feature importance and then served as inputs for machine learning models. Four models, namely support vector regression (SVR), multilayer perceptron (MLP), long short-term memory (LSTM), and temporal fusion transformer (TFT), were evaluated using root mean squared error (RMSE), coefficient of determination (R² score), and mean absolute error (MAE). SVR achieved the best generalization (RMSE = 0.293, R² = 0.854, MAE = 0.231), followed closely by MLP (RMSE = 0.288, R² = 0.859, MAE = 0.235). LSTM showed adequate learning ability but moderate testing performance, while TFT exhibited strong training accuracy but weak testing generalization. These results emphasize SVR’s robustness for seismic energy forecasting in data-limited conditions.

Research highlights

A global earthquake catalogue spanning over 500 years was used to construct an annual seismic energy release time series.

Empirical Mode Decomposition (EEMD) was applied to extract intrinsic mode functions (IMFs) capturing multi-scale seismic trends.

Machine learning models, including SVR, MLP, LSTM, and TFT, were developed for forecasting future energy release using IMF components.

Results demonstrate the feasibility of supervised learning for anticipating future global seismic energy release trends.