Abstract
Successful uncertainty quantification requires many reservoir models matching field production data and time lapse seismic. We have been able to automatically generate nearly 50 history matched models for a GOM oil field using the Neighbourhood Approximation stochastic sampling algorithm. This allows us to produce uncertainty forecasts for various production scenarios.
Geostatistical simulation constrained by well log, core, and seismic data is applied for building geological and reservoir models. The parameters of the geostatistical simulation (channel directions, channel dimensions, variagram parameters, etc.) are considered as uncertain parameters. Additionally, end points of relative permeability curves, dependencies of compressibility factors and permeability on effective stress, transmissibility multipliers across faults, and water aquifer size are varied in the history matching process.
A misfit function is selected to quantify history match of model results with field measurements of water cut, gas-oil ratio, and reservoir pressure in production wells in different moments of time. Trends and variances of the observation data are determined and incorporated in the objective function.
The position of the water-oil contact at the beginning of the field development and after three years of production is estimated from time lapse seismic. Differences in the water-oil contact positions determined from the reservoir simulation and time lapse seismic are quantified and incorporated in the objective function.
The reservoir model has been run nearly 2400 times in the history matching process. The Neighbourhood Algorithm is applied for the selection of the values of the history matching parameters in each run. The high quality of the history match is demonstrated,
An ensemble appraising procedure based on a Bayesian framework is used to determine probability distributions of the history matching parameters and to assign probabilities to the simulation models (runs). About 50 models with the highest probabilities (which cover 99% of the cumulative probability range) are selected for the predictions. A general tool has been developed for the definition of statistical parameters of production predictions (mean values, confidence interval, etc) and their changes in time. Uncertainty estimations for the base case predictions and several production scenarios are demonstrated.
