Cumulative oil production in flow unit-crossing wells estimated by multilayer perceptron networks

Abstract

Knowing the ultimate oil production in wells is a crucial point for reservoir planning and management to anticipate value for money. Commercial reservoir simulators are able to predict production curves with high confidence, but repetitive tasks in a few cases may spend a precious time of staff as well as require a large computational effort. Although artificial intelligence (AI) is providing an alternative path to the usual workflow, many commercial simulators lack robust AI algorithms. This work introduces a methodology based on a multilayer perceptron (MLP) neural network to predict the final cumulative oil production of a reservoir at vertical wells that cross hydraulic flow units (HFUs), which are volumes endowed with good flow attributes. Each well location is attached to special spots previously determined from clustering and calculation of maximum closeness centrality points (MaxCs) within a class of HFUs. The database is divided into training, validation, and testing sets organized after processing the UNISIM-I-D synthetic model, representative of the Namorado Field, Campos Basin, Brazil. The key rationale of this paper is to use the feature of MaxCs of being drivers for well placement as knowledge base to learn the production mechanisms of the oilfield. The outcomes are presented from two perspectives: an original MLP and its post-processed version. Both are compared with reservoir simulations carried out in CMG Imex^{copyright } and achieve reasonable agreement. The performance is measured by root-mean-squared error (RMSE) and mean absolute scaled error (MASE) both in original and post-processed versions. We show that average RMSE and MASE values near 0.07 and 14.00, respectively, are achieved without post-processing. With post-processing, gains of up to 43% are reported for the integral oil volume.