Perpustakaan Digital ITB

This paper presents a machine learning approach to generate a proxy model with high accuracy in predicting optimum well spacing of polymer flooding projects. The model used in this study is a reservoir with a 5-spot injection pattern mechanism of a polymer flood. The obtained proxy model can be used as a preliminary evaluation to determine the optimum well spacing injection scenario that can give maximum Profitability Index (PI) as the main objective function of this study. Understanding the physical phenomena and collecting historical project records of polymer floods are required to create an effective proxy model with a specific objective function and several parameter inputs. Once the parameterization range value is determined, the proxy model may be utilized practically. CMG-CMOST is used to generate 3655 experiments through Latin Hypercube Sampling (LHS) method to construct a reliable proxy model. Quality control is used to improve the proxy model quality by considering statistical knowledge and particular constraints. The proxy model developed with polynomial regression gives a quite good value of R-Square, 0.903, with the mean absolute error of 0.373. Another method constructed by CMG-CMOST is applied to increase the R-square value and decrease the error. The multilayer neural network method is applied with an architecture layers-neuron of 25-16, giving the best R-square of 0.975 and the mean absolute error of 0.217. To validate the optimization outcomes, the generated proxy model was used to evaluate the polymer flood project on a field scale, and it was successful in increasing the project's Profitability Index value.