An Ensemble Random Forest Model for Seismic Energy Forecast
Abstract. Seismic energy forecasting is critical for hazard preparedness, but current models have limits in accurately predicting seismic energy changes. This paper fills that gap by introducing a new ensemble random forest model designed specifically for seismic energy forecasting. Building on an existing paradigm, provided by Raghukanth et al. (2017), the global energy time series is decomposed into intrinsic mode functions (IMFs) using ensemble empirical mode decomposition for better representation. Following this approach, we split the data into stationary (IMF1) and non-stationary (sum of IMF2-IMF6) components for modeling. We acknowledge the inadequacy of intrinsic mode functions (IMFs) in capturing seismic energy dynamics, notably in anticipating the final values of the time series. To address this restriction, the yearly seismic energy time series is also fed along with the stationary and non-stationary parts as inputs to the developed models. Here, we employed Support Vector Machine (SVM), Random Forest (RF), Instance-Bases learning (IBk), Linear Regression (LR), and MultiLayer Perceptron (MLP) algorithms for the modelling. Furthermore, the five models discussed above were suitably employed in a novel regression-based ensemble random forest algorithm to arrive at the final predictions. The root mean square error (RMSE) obtained in the training and testing phases of the final model were 0.127 and 0.134, respectively. It was observed that the performance of the developed ensemble model was superior to those existing in literature (Raghukanth et al., 2017). Further, the developed algorithm was employed for the seismic energy prediction in the active Western Himalayan region for a comprehensively compiled catalogue and the mean forecasted seismic energy for year 2024 is 7.21 × 1014 J. This work is a pilot project that aims to create a forecast model for the release of seismic energy globally and further application at a regional level. The findings of our investigation demonstrate the possibility of the established method in the accurate seismological energy forecast, which can help with appropriate hazard preparedness.