Summary
Time-lapse (or 4D) seismic is increasingly being used as a qualitative description of reservoir behavior for management and decision-making purposes. When combined quantitatively with geological and flow modeling as part of history matching, improved predictions of reservoir production can be obtained. Here, we apply a method of multiple-model history matching based on simultaneous comparison of spatial data offered by seismic as well as individual well-production data. Using a petroelastic transform and suitable rescaling, forward-modeled simulations are converted into predictions of seismic impedance attributes and compared to observed data by calculation of a misfit. A similar approach is applied to dynamic well data. This approach improves on gradient-based methods by avoiding entrapment in local minima. We demonstrate the method by applying it to the UKCS Schiehallion reservoir, updating the operator's model. We consider a number of parameters to be uncertain. The reservoir's net to gross is initially updated to better match the observed baseline acoustic impedance derived from the RMS amplitudes of the migrated stack. We then history match simultaneously for permeability, fault transmissibility multipliers, and the petroelastic transform parameters. Our results show a good match to the observed seismic and well data with significant improvement to the base case.
Introduction
Reservoir management requires tools such as simulation models to predict asset behavior. History matching is often employed to alter these models so that they compare favorably to observed well rates and pressures. This well information is obtained at discrete locations and thus lacks the areal coverage necessary to accurately constrain dynamic reservoir parameters such as permeability and the location and effect of faults. Time-lapse seismic captures the effect of pressure and saturation on seismic impedance attributes, giving 2D maps or 3D volumes of the missing information. The process of seismic history matching attempts to overlap the benefits of both types of information to improve estimates of the reservoir model parameters.
We first present an automated multiple-model history-matching method that includes time-lapse seismic along with production data, based on an integrated workflow (Fig. 1). It improves on the classical approach, wherein the engineer manually adjusts parameters in the simulation model. Our method also improves on gradient-based methods, such as Steepest Descent, Gauss-Newton, and Levenberg-Marquardt algorithms (e.g., Lépine et al. 1999;Dong and Oliver 2003; Gosselin et al. 2003; Mezghani et al. 2004), which are good at finding local likelihood maxima but can fail to find the global maximum. Our method is also faster than stochastic methods such as genetic algorithms and simulated annealing, which often require more simulations and may have slower convergence rates. Finally, multiple models are generated, enabling posterior uncertainty analysis in a Bayesian framework (as in Stephen and MacBeth 2006a).
Ensemble Machine Learning Assisted Reservoir Characterization Using Field Production Data–An Offshore Field Case Study