Performance comparison of induced seismicity forecasting models with existing datasets

Within the workflow of Adaptive Traffic Light System, it is important to evaluate the performance of different induced seismicity forecasting models in order to properly weight the forecasts during seismic hazard calculation. In this respect, we propose a standardize test bench approach capable of comparing outputs’ models (in terms of seismicity rate) and their uncertainties in real time. We test this approach using different models that are trained using existing datasets from geothermal exploration campaigns. Notably, we use two statistical models that link injection volumetric rate to seismicity rate with the difference that a Bayesian approach (EM1_BH) additionally adds epistemic uncertainty to the aleatoric uncertainty introduced in a purely frequentist approach (EM1_MLE), one pressure-based seismicity model (HM0_CAPS) based on 1D analytical solution for linear pore-fluid diffusion and finally one hybrid 1D model that includes a physic-based module for linear and non-linear pore-fluid diffusion linked to a stochastic model for seismicity generation using a seed approach (HM0_SEED and HM1_SEED). By using these different models and their uncertainties in our numerical investigations, we show the robustness of the proposed testbench approach.