The objective of this project is to explore cutting-edge sequence-based machine learning models commonly used in language processing to reproduce a multi-porosity reservoir simulator. The proposed method integrates advanced techniques to significantly reduce the numerical simulation time and improve the decision-making process for Huff and Puff (H-n-P) gas injection optimization in shale reservoirs. The proposed approach follows three crucial steps to predict an output sequence given an input sequence: 1) the simulation results should be validated against actual data, 2) train and validate a machine learning model using simulation results from either commercial or in-house numerical simulators, 3) exhaustive exploration of hyperparameter tuning and selection of machine learning techniques, such as sequence-to-sequence (Seq2Seq), Luong attention and ConvLSTM. The proxy model considers as input variables well control parameters such as injection and production periods, number of cycles and gas injection rates to estimate the proxy model results.

The multi-porosity proxy reservoir simulation model is a complementary tool that integrates numerical simulation and data-driven techniques. Although tuning the model typically demands significant time, it can speed up the simulation time up to 20,000X allowing for generating hundreds or even thousands of scenarios at the expense of accepting a reduction in the accuracy of the results in a matter of minutes. One of the most notable findings is that considering a small training dataset, the proxy model can reproduce the capabilities for predicting oil production in complex low and ultra-low permeability reservoirs with significantly reduced error, relative to the multi-porosity reservoir simulator. Finally, the possibility of reproducing a considerable number of scenarios in minutes opens the door to exploring different well control configurations such as injection and production periods, number of cycles and gas injection rates. The novelty of the proxy multi-porosity reservoir simulator is to notably accelerate the numerical simulation time by using techniques capable of solving sequence learning problems in which the output is dependent on previous outputs.

You can access this article if you purchase or spend a download.