An Application of Feed Forward Neural Network as Nonlinear Proxies for Use During the History Matching Phase

Abstract

Abstract Reservoir simulation is an important tool used in the industry for reservoir management. While developing a field, a reservoir simulation model is used as a decision tool to select the best development scheme and also to forecast the oil, gas, and water production expected for the field. Uncertainties are much higher at the early phases and, when production data are gathered during the field development phase, most of the time the initial reservoir simulation model needs to be reviewed once the field observed data is not as the same as the predicted by the model. Some of these uncertainties of these input parameters are related to the reservoir rock reservoir heterogeneities. History matching techniques are used by reservoir engineers to mitigate/minimize the difference between the observed field data and the predicted data and thus assessing the uncertainties. When reservoir models become too big in terms of number of cells and features, the elapsed simulation time increases very much, making the history matching process very cumbersome and, in some cases, very difficult to achieve in an acceptable time. Parallel processing features of some commercial simulators can perform lots of simulation runs at the same time but cannot address and cannot solve the problem in a proper way. This paper presents an alternative proposition to speed up the history matching process: the application of feed-forward neural networks as nonlinear proxies of reservoir simulation. Neural networks can map the response surface in multidimensional spaces of a reservoir model (i.e. water production, bottom hole pressure etc.) or of an objective function with few number of simulations. The mapped response is then used as a substitute of reservoir simulation runs during the history matching process. The focus of this work is to shown the steps of choosing the best number of hidden layers, the neurons and the training method. An application case is presented using the workflow presented is this work and showing the validity of the proposed methodology for this complex nonlinear problem.