Using machine learning for model benchmarking and forecasting of depletion-induced seismicity in the Groningen gas field

Abstract

The Groningen gas field in the Netherlands is experiencing induced seismicity as a result of ongoing depletion. The physical mechanisms that control seismicity have been studied through rock mechanical experiments and combined physical-statistical models to support development of a framework to forecast induced-seismicity risks. To investigate whether machine learning techniques such as Random Forests and Support Vector Machines bring new insights into forecasts of induced seismicity rates in space and time, a pipeline is designed that extends time-series analysis methods to a spatiotemporal framework with a factorial setup, which allows probing a large parameter space of plausible modelling assumptions, followed by a statistical meta-analysis to account for the intrinsic uncertainties in subsurface data and to ensure statistical significance and robustness of results. The pipeline includes model validation using e.g. likelihood ratio tests against average depletion thickness and strain thickness baselines to establish whether the models have statistically significant forecasting power. The methodology is applied to forecast seismicity for two distinctly different gas production scenarios. Results show that seismicity forecasts generated using Support Vector Machines significantly outperform beforementioned baselines. Forecasts from the method hint at decreasing seismicity rates within the next 5 years, in a conservative production scenario, and no such decrease in a higher depletion scenario, although due to the small effective sample size no statistically solid statement of this kind can be made. The presented approach can be used to make forecasts beyond the investigated 5-years period, although this requires addition of limited physics-based constraints to avoid unphysical forecasts.