Artificial Neural Network Modeling of Well Performance in the Garrington Field, Cardium Formation

Abstract

A number of studies have reported the use of artificial neural networks (ANNs) to predict tight formation well performance. ANNs are attractive because they do not require pre-conceived models (are “data-driven”), can accommodate numerous inputs, and allow for nonlinear relationships. There are reports of using ANNs in ‘factory mode’ to aid well and stimulation design where fast development of areas precludes devoting the time and effort to individually design each well. The analysis and experience of this study shows, however, that ANNs have limitations which can go unrecognized and lead to faulty predictions: 1) We find that investigators may not have adequately tested the ANN by overlooking the performance for the testing and validation tests, and only reporting the “R” values (correlation coefficients) for the training results. 2) There is a large variability between ANNs trained using the same dataset. We find that, even when we select only those ANNs which give R > 0.85, there can be significant discrepancies between predicted and actual performance. 3) ANN models which exclude the operator may give poor results and make some variables appear more important than they actually are. 4) Among other explanatory variables, the early time linear flow parameters are very important. This thesis illustrates this work using data from the tight oil Garrington Field (Cardium Formation). It makes recommendations to ANN workflows which will guide practitioners in the appropriate development, testing, and application of ANNs in this important topic.