Optimization and control of reservoir models using system identification and machine learning tools

Abstract

This dissertation develops a methodology to identify a dynamical system modeling an oil and gas reservoir, subject to production controls such as water injection rate and liquid production rate. The overall objectives are to improve production forecasts and decision making processes regarding the development of the field, such as future control strategies and “what-if” analyses considering different scenarios. The classical history matching approach uses numerical simulation and tuning of geological parameters. In contrast, this dissertation proposes the use of a system identification approach to build two proxy models, one based on the input-output approach and the other on a state-space approach, both utilizing data that comes from a simulator used in industry. In accordance with the parsimony principle, simpler polynomial model structures such as ARX and ARMAX are used for the input-output model. The linear state space model uses states coming from model simulation as its data for identification, and is subjected to model reduction using the proper orthogonal decomposition (POD) method. This linear state space reduced order proxy model is then used to formulate an optimal control problem, solved by transcription into a sequence of linear programs using a trust region algorithm, maximizing Net Present Value, which is an objective function representing the overall economic performance of the production process. Additional significant contributions, developed in the course of this dissertation, include a fast method for uncertainty estimation, analysis of RLS coefficient adaptation to get physical insights into the correlation between producer and injector wells, as well as insights into model selection and incorporation of prior knowledge.