Development of artificial neural networks for steam assisted gravity drainage (SAGD) recovery method in heavy oil reservoirs

Open Access
Author:
Sengel, Ayhan
Graduate Program:
Petroleum and Natural Gas Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
July 29, 2013
Committee Members:
  • Turgay Ertekin, Thesis Advisor
Keywords:
  • Artificial neural network; sagd; steam; horizontal wells; gravit
Abstract:
As no alternative energy source has been introduced to efficiently replace fossil fuels yet, the demand for oil and gas is still increasing in the world. Conventional hydrocarbon reservoirs have been depleted rapidly to meet the demand; in doing so, the amount of conventional resources has declined. This has led to the interest of methods to enhance hydrocarbon recovery in unconventional resources such as heavy oil, bitumen and oil shale. As a result of its high viscosity, economic recovery of heavy oil and bitumen is a great challenge. Steam Assisted Gravity Drainage (SAGD) is a commercial in-situ recovery technology used to reduce oil viscosity by increasing the temperature of the reservoir to enhance the recovery of heavy oil and bitumen. Two parallel horizontal wells are employed in the SAGD process. The upper horizontal well is employed for continuous steam injection into the reservoir, while the lower horizontal well is used to produce reservoir fluids. Numerical simulation of SAGD is a complex task. Instead of struggling with high complex problems, simple models are used to substitute reservoir simulations over a given range of input parameters. Usage of neural-network based proxy models for the solution of non-linear relationships has been increasingly popular in recent years in the oil and gas industry. Artificial neural network (ANN) methodology is used to avoid excessive time consumption of computer simulations. The objective of this study is to develop neural-network based proxy models that can provide instant and reasonably accurate preliminary estimations for SAGD applications. The variables used in this study include reservoir rock/fluid properties, such as reservoir thickness, porosity, horizontal permeability, horizontal-vertical permeability ratio, initial reservoir pressure, reservoir temperature, rock thermal conductivity, initial oil saturation and oil density, together with operational variables including vertical spacing, spacing between the producer and the base of the reservoir, inter-well (well pattern) spacing, sub-cooling temperature, steam quality, well length, maximum injector bottom-hole pressure and minimum producer bottom-hole pressure. The present study aims at developing neural-network based proxies for SAGD recovery method in heavy oil reservoirs. A two-stage approach was used. In stage I, a total of 904 reservoir specific samples have been trained. The forward-looking artificial neural network is used to predict performance indicators such as cumulative oil production, cumulative steam oil ratio and cumulative water production profiles over a period of 10 years for a given set of operational (design) parameters. On the other hand, the inverse-looking artificial neural network is used to predict operational parameters for a given set of desired 10-year cumulative oil production and cumulative steam-oil ratio profiles. In stage II, reservoir properties are varied as well. ANN methodology is extended to a range of homogeneous reservoirs. A total of 1590 samples were generated. This stage provides a forward-looking artificial neural network that predicts performance indicators over a period of 10 years and two inverse-looking artificial neural networks: one that predicts operational parameters and another that predicts reservoir properties. CMG CMOST Studio automation tool was used to generate SAGD numerical simulation samples over a given input data range. Numerical simulations were performed using the CMG STARS thermal simulator (version 2011.10). Numerical simulations were utilized to feed the training of the artificial neural network using MATLAB neural network toolbox (version R2011a). The results of this study show that artificial neural networks are able to recognize complex relationships between input data and corresponding output data for SAGD simulations. Therefore, ANNs are suitable tools for SAGD forecasting and analysis.