Eventos Anais de eventos
ENCIT 2022
19th Brazilian Congress of Thermal Sciences and Engineering
Estimation of well production flow rate using Recursive Neural Networks
Submission Author:
Sergio Ribeiro , RJ
Co-Authors:
Rafael Dias, Sergio Ribeiro, Marcio CARVALHO
Presenter: Sergio Ribeiro
doi://10.26678/ABCM.ENCIT2022.CIT22-0233
Abstract
Reservoir characterization is a very important task in the oil and gas industry. The common approach is to stop production to perform well tests. During the testing period, the production well is either controlled at a constant flow rate (Drawdown) or it is closed (Buildup). The test can take a few hours up to a few days to generate all data needed to estimate the reservoir parameters based on the transient response of the well-reservoir system. In most of the recently built wells, permanent downhole gauges (PDGs) are being installed, enabling pressure and temperature data collection during production. Most of these sensors do not include flowrate measurements. The ability to estimate flow rate of producing wells based on the transient response of pressure and temperature during production could speed up the estimation of evolution of reservoir parameters and the overall management of the oil field. This estimation can be done by coupling the solution of an inverse problem and a reservoir simulator. This alternative is computationally very expensive and not practical in most applications. An alternative, proposed by Tian et al (2018), is to use deep neural networks to estimate production flow rate and perform reservoir properties estimation using the PDG data. In this work, we use a Recursive Neural Network (RNN) to study the behavior of a reservoir with known permeability. The pressure and temperature profiles are generated by a reservoir flow simulator, with a prescribed flow rate time evolution input. Part of this dataset is passed as training data to the RNN. Therefore, the trained RNN flow rate predictions for an unseen dataset are then compared with the simulated ground truth. Finally, a sensitivity analysis is performed changing the sliding window size, RNN architecture and training parameters. The results obtained for data generated with the simulator can also be compared with the literature results of networks trained with real wells pressure test data.
Keywords
machine learning, Well Testing, Prediction methods, Production planning
