Time-lapse seismic is an important subject in the Petroleum Industry. Going beyond the qualitative interpretation of the seismic data, including quantitative information in the flow simulation model updating process, is a highly desirable goal. 4D seismic data give information far away from the wells, potentially allowing much richer parameterizations of the reservoir model in the history matching process. Since these parameterizations tend to be described by a large number of parameters, efficient algorithms are needed to tackle these problems.
This paper describes some efforts to integrate time-lapse seismic attributes into a derivative-based assisted history matching tool developed in a previous project 1. The implementation of the seismic attributes derivatives using the forward and the adjoint method into a commercial reservoir flow simulator is described. The calculated derivatives are used in an optimization algorithm based on a trust-region Quasi-Newton method to minimize the mismatch between observed and simulated data from production and seismic.
Good results were obtained in several synthetic cases adjusting seismic and production data.
The history matching process is usually one of the most complex parts of a reservoir simulation study. Basically, the history match process tries to modify the flow simulation model input data (permeability, porosity, fault transmissibility, etc.) in order to make the model fit observed production data 2. The current industry practice is still, in most cases, a manual procedure of trial and error that requires a lot of experience and knowledge from the geoscientists involved in the study. The process is also inherently non-unique and can have several solutions. Nevertheless, the history matching procedure is essential because it ensures some legitimacy to the flow simulation model.
The development of computer hardware and software allowed the appearance of several assisted history matching tools 1,3,4,5,6. These tools are based on the minimization of a least-squares objective function (OF) that quantifies the misfit between the simulated and observed data. The classical approach to history matching a reservoir model uses the production data, oil, water and gas rates and pressure measured at wells, as observed data to be fitted by the simulator.
With the improvement of seismic technology, a new kind of data emerged in the oil industry. This new data consists of a series of seismic images of the reservoir taken in different moments of the production life of a petroleum field 2, usually called as time-lapse or 4D seismic. Basically, the 4D seismic may help to identify areas of bypassed oil for infill-drilling wells or locate water flooded areas, avoiding the drilling of low-performance wells in swept regions, to map gas cap areas, to identify compartmented regions in the reservoir and flow barriers. In this sense, the 4D seismic data contains relevant information of the fluid and pressure changes in the region away from the wells. Therefore, this kind of data has started to be used into history matching problems, initially in a qualitative manner, for instance, by manipulating reservoir parameters to match observed flow paths.
Recently quantitative information from 4D seismic data has started to be incorporated into the history matching process. Huang et al.7,8 have presented a history matching procedure that includes production data and 4D seismic amplitudes. An application to a dry-gas reservoir model in Gulf of Mexico has shown that this approach can improve the reliability of the model predictions. The process was based on a stochastic search procedure to minimize the misfit between the simulated and observed data. Waggoner et al.9 have applied a similar approach in a gas condensate reservoir using acoustic impedances with production data in the OF. Stephen et al.10,11 have presented a method to generate multiple history matched models based on production and seismic data. In this work the misfit between the observed and simulated data was used to update the model probabilities in a Bayesian framework.