Abstract

Seismicity induced by fluid injection (e.g. enhanced geothermal systems, waste water
disposal, CO2 storage) remains a significant risk to communities and industry as it may
cause structural damages and economic losses. In some cases, this can even lead to the
immediate shutdown of the project, as observed at the Pohang geothermal site in South
Korea (2017), after inducing a magnitude 5.5 earthquake. Consequently, it is crucial
to develop new strategies to anticipate and mitigate the seismic risk posed by these
operations. In this work, we present an innovative machine learning-based approach to
forecast the seismicity induced by fluid injection into the ground. It leverages the time
series of various injection parameters such as injected volume, flow rate and wellhead
pressure, to capture their relationship with the seismicity rate using machine learning.
Once our model is trained, it is then possible to forecast the future number of induced
earthquakes, on a fixed time scale of interest. We apply this strategy to two case studies
having caused induced seismicity at different spatial and temporal scales. Firstly, we
investigate the relationship between the massive volume of wastewater injected in Okla-
homa (USA) since 2006 and the sharp increase in seismicity. Capturing this relationship
with a linear regression, it is then possible to forecast the seismicity associated to any
future hypothetical injection scenario. Secondly, we focus on the seismicity induced
during the stimulation phase of the Soultz-Sous-Forets (France) Enhanced Geothermal
System (EGS) in 2005. Using a gradient boosting model, we predict the seismicity
over the next 24 hours, relying solely on the injection parameters. This work paves
the way for real-time application, as the prediction of microseismicity within ongoing
geothermal operations at the Rittershoffen EGS site, located just 6 kilometers away
from Soultz-Sous-Forets.