History Matching: Is it Necessary to Optimize?

Abstract

Abstract The oil industry has recently started to deal with probabilistic approach. Risk or uncertainty analysis have become part of the petroleum engineer's job. A set of curves with the associated probability instead of one deterministic curve is provided by the reservoir engineers. In order to use reliable curves, they shall have a history matched model. Assisted History Matching usually uses optimization processes. The aim of the optimization is to find the minimum of an objective function that represents the quality of the model. In this way, one can find the best model. The keyword is exactly "best". Why to make so much effort to find the best if we know that it is still far from the truth. Indeed, the concept of "best" is not suitable for the probabilistic approach. This work discusses a functional history matching approach where an optimization process is no longer necessary. The functional history matching approach establishes that we have to look for a set of models that is above a level of quality according to the reservoir engineers. The method is quite simple. Among all possible models, we select those that have an objective function value under a pre-defined value. In this approach the discussion lies not in the optimization issues like local minimum, convergence, and rapidity, but in how the quality of the model is measured. The objective function that usually measures the quality must be very well defined. Not only to better take into account the historical data but also to be suitable to the purpose of the study. Infill drilling and new secondary recovery systems would probably require different objective functions. This work discusses the functional history matching approach coupled with uncertainty analysis. Usually very costly in terms of numerical simulations, uncertainty analysis can be done in this approach with simplified models (proxys). Different proxys were used - Surface Response Modeling (improved or not) and Artificial Neural Network. A simple synthetic case (PUNQ), and a real complex case (Brazilian onshore field) were used to illustrate the functional approach. Introduction The main motivation for this work is to discuss the history matching process in the way it is being applied in most cases. Despite many companies using probabilistic approaches in their studies, a huge effort has been made to improve their models using different optimization processes. The question is: is it necessary? The former "Automatic History Matching" has given rise to "Assisted History Matching" because automatic processes have showed itself very risky. The reservoir engineer could loose control of the process and the final model could be unreliable. An assisted process is then used to guarantee that the reservoir engineer keeps all variables under control. Nevertheless, the concept of the "best model" persists. "Best model" depends on the goal of the study: existing well production forecast and infill drilling project require different quality levels of the model. Additionally, the best model becomes "old model" very rapidly. New data must be incorporated as soon as possible, and, sometimes, important changes have to be made. As a rule, this must be done quickly, and the reservoir engineer hardly keeps the model accurate. The probabilistic approach has risen because companies need to manage risks and project flexibility and to deal with the multiplicity of scenarios. The range of possible models can be wide, and, if production data are available, they have to be taken into account in order to have more reliable models. This process is called "uncertainty analysis with history matching". In this work, a functional history matching approach is discussed and it could be called "history matching with uncertainty analysis". The expression inversion is not just a semantic issue, but a change in the way history matching can be seen. In this functional history matching approach a best model is no longer sought after. "Best" means unique and it doesn't match with the companies needs. A unique "best model" may sometimes be useful but it is certainly not sufficient.