Statfjord study demonstrates use of neural network to predict porosity and water saturation from time-lapse seismic

With an estimated STOIIP of more than 1 billion m3, Statfjord is the largest oil discovery in the North Sea to date. Structurally the field is a tilted fault block with Jurassic beds dipping westward at 6 to 8° and truncated on the faulted and eroded East flank (Kirk, 1980 and Fig. 1). The field is divided into three main reservoir units, which are in order of decreasing importance: Brent, Statfjord and Dunlin. Since its discovery in 1974 almost 200 wells have been drilled. The cumulative production from the start of production in 1979 until the end of 1997 was 550 million m3. This represents approx. 55% of the initial oil in place and 83% of the official recoverable reserves of 662 million m3. All reservoirs have been partly drained so far during the field's history. In the early years, produced gas was reinjected into the Statfjord reservoir at an up dip position, while the Brent reservoir was depleted until pressure maintenance by down flank water injection was established in 1986. This resulted in extensive gas and water breakthroughs in the production lines and a change in drainage strategy. In the last decade the strategy has been based on in-fill drilling to produce by-passed oil and remaining oil in structural traps and recently in combination with gas and water injection to mobilize remaining oil. To find unswept or by-passed oil three consecutive 3D seismic surveys, acquired in 1979, 1991 and 1997 respectively, have been the basis for seismic monitoring analysis. In this case study, only the latter two surveys were used. The aim of the study was to predict porosity and water saturation at the respective acquisition dates. For each survey five seismic volumes were available: mid- and far-angle reflectivity, mid- and far-angle elastic impedance, and acoustic impedance. The partial stacked cubes have similar fold and contain angles around 15 and 25°, respectively. The inversion method used was a global search through a simulated annealing scheme using a constant wavelet. Some 130 wells were used, each with an extensive suite of measured logs and 1991 and 1997 timeequivalent logs. The reservoir simulator and a modified Gassmann fluid replacement algorithm were used to compute the time-equivalent logs. For most wells a neural network predicted a measured shear sonic log from measured sonic, density and gamma ray logs.