CONTINUOUS CO2 INJECTION DESIGN IN NATURALLY FRACTURED RESERVOIRS USING NEURAL NETWORK BASED PROXY MODELS
Open Access
- Author:
- Hamam, Hassan
- Graduate Program:
- Petroleum and Mineral Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 13, 2016
- Committee Members:
- Turgay Ertekin, Dissertation Advisor/Co-Advisor
- Keywords:
- ANN
Artificial Neural Networks
CO2 Injection
CO2
Continuous CO2 Injection
Carbon Dioxide injection
CO2 EOR
ANN CO2
CO2 Artificial Neural Networks - Abstract:
- More than 60% of the original oil in place (OOIP) is left in the ground after the primary and secondary recovery processes. With the introduction of enhanced oil recovery (EOR), that number goes down to about 40% of the OOIP. Carbon dioxide (CO2) injection is one of the most effective EOR methods in naturally fractured reservoirs. The fracture network provides a faster means for fluid flow due to its high conductivity but it is also the cause of premature breakthrough of the injected fluids. However, if employed efficiently, fractures can help push the injected CO2 to the reservoir boundaries so that a large portion of the reservoir fluid interacts with the injected CO2. Zones swept by miscible CO2 reported the lowest residual oil saturation. Continuous CO2 injection is becoming more and more preferred to the popular cyclic pressure pulsing. Continuous CO2 injection has no down time and could potentially provide better CO2 interaction with the reservoir fluid which provides a higher recovery. In this research, artificial neural networks (ANNs) are used to construct robust proxy models with highly predictive capabilities for naturally fractured reservoirs undergoing continuous CO2 injection. The main purpose of this research is to shed more light and understanding on continuous CO2 injection in naturally fractured reservoirs and provide a tool that empowers engineers to make decisions on the fly while evaluating uncertainty and mitigating risk rather than wait months or years to do so. In light of the above, various ANN designs and configurations undergo development and evolution to ultimately be able to provide valuable insights regarding reservoir performance, history matching, and injection design for naturally fractured reservoirs undergoing CO2 injection. Initial ANN designs targeted specific reservoirs using specific fluid compositions from the literature. The designed ANNs were able to provide predictions with a low degree of error. ANN designs went over many complex adjustments, variations, and enhancements until final configurations were reached. The final ANN designs developed in this research surpass previously developed ANNs in similar projects with its capability to handle a huge range of reservoir properties, relative permeability, capillary pressure, and fluid compositions under uncertainty. The reservoir simulation model used in this research is a two-well, two-layer, miscible compositional simulation model working in a dual-porosity system. Critical parameters affected the accuracy and predictability of the ANN designs and they were an essential part of the final ANN configurations. The parameters that a major effect on continuous CO2 injection are reservoir fluid composition, fracture permeability, well spacing, bottomhole flowing pressure (BHFP), thickness, and CO2 injection amount under miscible conditions had the highest impact on recovered oil. The final ANN designs were encompassed inside a graphical user interface that equipped the ANN with uncertainty evaluation capabilities. The ease to use nature of the GUI allows anyone to use the developed ANNs in this research, as well as provide a simple intuitive interface to manipulate input data, run simultaneous sensitivity and uncertainty analysis. The developed ANNs in this research bring us a step closer to achieving real-time simulation for naturally fractured reservoirs undergoing CO2 injection. The correlations embedded in the ANNs were able to overcome reservoir fluid, relative permeability, and capillary pressure limitations that existed in the previous ANN studies.