Building Robust Models to Forecast the Induced Seismicity Related to Geothermal Reservoir Enhancement

Near real-time seismicity forecasting is critically important for decision making during the reservoir creation phase of enhanced geothermal systems (EGS). To compute within a probabilistic framework the seismic hazard and risk posed by induced seismicity, it is necessary to calibrate robust forecast models, quantitatively assess their reliability, and capture the uncertainty of the process. Here, we test the performance of three model classes, each forecasting the induced seismicity observed during the failed 2006-2007 EGS reservoir creation underneath Basel, Switzerland. Each model class represents a logic tree branch that captures the epistemic uncertainty of the process; the aleatory variability is accounted for by bootstrapping of the observed seismicity. We calibrate each model using a range of updating strategies and quantitatively test the performances of the models using likelihood tests. We then define a combined model based on Akaike weights and show that the combined model outperforms, in our pseudoprospective testing environment, all of the individual models. We also show that seismicity recorded during the initial one to three days of seismicity, several days before the onset of felt seismicity, is sufficient for deciding with confidence that the prescribed injection regime will eventually lead to unacceptably high levels of seismicity.