Integrating New Data in Reservoir Forecasting Without Building New Models

In most hydrocarbon reservoir development projects, geological models are fully rebuilt on a regular basis to integrate new data, in particular observations from new wells, for up-to-date forecasts. Not only this common practice is very time consuming as rebuilding models can take weeks or even months, but it also leads to major, hard-to-justify, fluctuations in reservoir volume or flow performance forecasts, especially when the modeling staff changes or a new modeling technology, workflow, or software is adopted. Rationalizing the geological model updating process is required to provide stable and reliable forecasting and make timely, well-informed, reservoir management decisions. This paper presents an innovative methodology to quickly reassess model forecasts, such as reservoir oil-in-place or oil recovery, without rebuilding any geological models provided that the new data observations are reasonably consistent with the current models. The proposed methodology uses a Bayesian framework whereby the multivariate probability joint distribution of new data predictions and forecast variables needs to be modeled. Assuming that this joint distribution is multi-Gaussian, the first step consists in computing proxies, e.g., response surfaces using experimental design, to estimate from the set of current geological models the distribution (mean and variance) of new data predictions and forecast variables as a function of the input modeling parameters (e.g., property variograms or training images, trends, histograms). Because the model stochasticity (i.e., spatial uncertainty away from wells) typically entails significant uncertainty in the prediction of new local data observations, computing the previous proxies requires generating multiple stochastic realizations for each combination of input modeling parameters. Then, using those proxies and Monte Carlo simulation, the full multivariate probability joint distribution of new data predictions and forecast variables is estimated. Plugging the actual new data values into that joint distribution finally provides new updated probabilistic distributions of the forecast variables. This new methodology is illustrated on a synthetic case study. In addition to quickly reassess reservoir volume and flow performance predictions, this new approach can be used to select new data observation types and impact maps to assess potential well locations that would optimally reduce forecasting uncertainties.