Geothermal systems are drawing large attention worldwide as an alternative source of energy. Although geothermal energy is beneficial, field operations can produce induced seismicity whose effects can range from light and unfelt to severe damaging. In a recent paper by Convertito et al. (2012), we have investigated the effect of time-dependent seismicity parameters on seismic hazard from induced seismicity. The analysis considered the time-variation of the b-value of the Gutenberg-Richter relationship and the seismicity rate, and assumed a non-homogeneous Poisson model to solve the hazard integral. The procedure was tested in The Geysers geothermal area in Northern California where commercial exploitation has started in the 1960s. The analyzed dataset consists of earthquakes recorded during the period 2007 trough 2010 by the LBNL Geysers/Calpine network. To test the reliability of the analysis, we applied a simple forecasting procedure which compares the estimated hazard values in terms of ground-motion values having fixed probability of exceedance and the observed ground-motion values. The procedure is feasible for monitoring purposes and for calibrating the production/extraction rate to avoid adverse consequences. However, one of the main assumptions we made concern the fact that both median predictions and standard deviation of the ground-motion prediction equation (GMPE) are stationary. Particularly for geothermal areas where the number of recorded earthquakes can rapidly change with time, we want to investigate how a variation of the coefficients of the used GMPE and of the standard deviation influences the hazard estimates. Basically, we hypothesize that the physical-mechanical properties of a highly fractured medium which is continuously perturbed by field operations can produce variations of both source and medium properties that cannot be captured by a stationary GMPE. We assume a standard GMPE which accounts for the main effects which modify the scaling of the peak-ground motion parameters (e.g., magnitude, geometrical spreading and anelastic attenuation). Moreover, we consider both the inter-event and intra-event components of the standard deviation. For comparison, we use the same dataset analyzed by Convertito et al. (2012), and for successive time windows we perform the regression analysis to infer the time-dependent coefficients of the GMPE. After having tested the statistical significance of the new coefficients and having verified a reduction in the total standard deviation, we introduce the new model in the hazard integral. Hazard maps and site-specific analyses in terms of a uniform hazard spectrum are used to compare the new results with those obtained in our previous study to investigate which coefficients and which components of the total standard deviation do really matter for refining seismic hazard estimates for induced seismicity.