STUDY OF THE INFLUENCE OF TRAINING DATA SET IN ARTIFICIAL NEURAL NETWORK APPLIED TO THE HISTORY MATCHING PROCESS

History matching is an inverse problem designed to find a combination of reservoir parameters that min i ize an objective function that represents the matching quality by measuring the difference between the sim ulated and observed data. Manual process is usually a time con suming and tedious task. An interesting alternative is the assisted history matching (AHM) that consists in the automat ion of part of the process. However, AHM normally d emands many simulations. Proxy models, generated from Arti ficial Neural Networks (ANN) for example, can be us ed in the process in order to reduce the number of simulation s. Artificial Neural Networks are becoming increasi ngly popular in the oil and gas industry. One of the fundamental as pect in the study of ANN is the definition of the training data set. This is a difficult task and is it strongly depende t on the problem. The difficult increases for prob lems with high number of variables and high nonlinearities. The ma in idea of present work is to analyze the influence of training data set to generate proxy models to represent the simul ator in the history matching process. Backpropagati on multiple-layer networks were trained using date set generated thro ug Latin Hypercube and Box-Behnken design techniqu e. Two different size (number of points) of the data set w as also tested and compared. The proxy models gener at d by the trained networks were used in the optimization proc ess with genetic algorithm (GA). The better solutio ns found by the GA were tested with the reservoir simulator to vali d te the results. A realistic reservoir model with eight producers and seven water injector wells and 16 uncertain paramet ers was used in this study. Rio Oil & Gas Expo and Conference 2010