Multi-objective optimisation metrics for combining seismic and production data in automated reservoir history matching
Chime, Onwunyili Christian
MetadataShow full item record
Information from the time-lapse (4D) seismic data can be integrated with those from producing wells to calibrate reservoir models. 4D seismic data provides valuable information at high spatial resolution while the information provided by the production data are at high temporal resolution. However, combining the two data sources can be challenging as they are often conflicting. In addition, information from production wells themselves are often correlated and can also be conflicting especially in reservoirs of complex geology. This study will examine alternative approaches to integrating data of different sources in the automatic history matching loop. The study will focus on using multiple-objective methods in history matching to identify those that are most appropriate for the data available. The problem of identifying suitable metrics for comparing data is investigated in the context of data assimilation, formulation of objective functions, optimisation methods and parameterisation scheme. Traditional data assimilation based on global misfit functions or weighted multi-objective functions create bias which result in predictions from some areas of the model having a good fit to the data and others having very poor fit. The key to rectifying the bias was found in the approaches proposed in this study which are based on the concept of dominance. A new set of algorithms called the Dynamic Screening of Fronts in Multiobjective Optimisation (DSFMO) has been developed which enables the handling of many objectives in multi-objective fashion. With DSFMO approach, several options for selecting models for next iteration are studied and their performance appraised using different analytical functions of many objectives and parameters. The proposed approaches are also tested and validated by applying them to some synthetic reservoir models. DSFMO is then implemented in resolving the problem of many conflicting objectives in the seismic and production history matching of the Statoil Norne Field. Compared to the traditional stochastic approaches, results show that DSFMO yield better data-fitting models that reflect the uncertainty in model predictions. We also investigated the use of experimental design techniques in calibrating proxy models and suggested ways of improving the quality of proxy models in history matching. We thereafter proposed a proxy-based approach for model appraisal and uncertainty assessment in Bayesian context. We found that Markov Chain Monte Carlo resampling with the proxy model takes minutes instead of hours.