Measuring the Value of Time-Lapse (4D) Seismic as Part of History Matching in the Schiehallion UKCS Field

In seismic history matching we use production data from wells and time-lapse (4D) seismic to constrain simulation models so that they better represent reservoir properties and behaviour. Together, these data types reduce the non-uniqueness of the problem, and therefore reduce the uncertainty of both the reservoir description and also the estimation of future behaviour. The more constraints we have, however, the harder it is to find the best models and more simulations may be required to search the parameter space. This leads to increasing computing costs, which must be balanced against the reduction in model uncertainty.<br> <br>We have developed a method of performing a cost:benefit analysis of including extra data and simulations in the history matching process. We use a Neighbourhood Algorithm to sample the parameter space and work in a Bayesian framework to determine model probabilities. After history matching, we then resample the posterior probability density to estimate parameter uncertainties. In addition, the parameter sampling has a density roughly in proportion to the probability distribution of the models. With this property of our method and with sufficient models, we then determine the most likely model outcome and its uncertainty. This enables calculation of expected saturation and pressure distributions at the time the data was measured and into the future. This is beneficial for reservoir management, particularly for identifying unswept areas.<br><br>We apply our method to a UKCS field and analyse how the uncertainty changes in response to adding the seismic data to the history match. We also analyse the change in uncertainty as a function of the number of simulations carried out. We identify an optimum number of models that are required before we enter the domain of diminishing returns. We confirm that seismic is important if we wish to describe the reservoir some distance from production wells. We also find that some parameters may be determined more quickly than others, depending on their location relative to the data being used. <br>