New Parallel Algorithm for Multiple History-Matched Realizations

Abstract

Abstract Correct prediction of reservoir performance is an important issue for future reservoir development and exploration planning. This leads to the problem of accurate estimation of the rock and fluid properties used in construction of reservoir model. Usually, prior reservoir models are built by means of geostatistical algorithms and static information obtained from different sources and then adjusted for better production history matching. However, in some cases, static data can be very sparse and insufficient or misleading, for meaningful prior model construction. Under this scenario, adjusting of prior model would lead to a local minimum in automatic history matching algorithms and hence we will obtain results which may not be meaningful. This paper presents a new methodology for Parallel Production data Processing. It assumes limited prior knowledge about static information and emphasizes the importance of production data. This methodology begins with several initial models which represent large amount of uncertainty in prior models. The method goes through a process of cyclic parallel adjustment of multiple reservoir models and reliable information gathering for further models improvements. The algorithm seeks common trends developed in the prior models as a result of partial incorporation of production data. Every generation of multiple models is adjusted by means of gradient-based automatic history matching technique. By using common trends gathered from prior models, new set of solutions (alternate realizations) are proposed. Generalized Changes Map (GCM) is the main source of information in this method. It provides the trend of common changes for all the models and serves as a basis of reliable parameters and spatial relationship selection. GCM captures two types of information: maximum changes occurred in prior model due to integration of production data, and convergence of values towards a common value in multiple models. These parameters form Generalized Distribution Map (GDM) and used as a conditional data for construction of a new generation of models. Procedures of models adjustment, parameters selection, and models construction are continued till desired level of convergence in objective function is reached. The final product of the process is the multiple realizations which are history matched. The method is validated using some synthetic cases as well as PUNQ S3 model.