Decision-making framework to optimize the waterflooding process in an oilfield using reduced-order models and machine learning

In the face of increasing global energy demand and the need for energy transitions, improved decision-making processes in the oil and gas industry are essential. Waterflooding is a successful method for enhancing oil recovery. Numerical reservoir simulation software are essential tools for evaluatin...

Full description

Autores:
Rodríguez Castelblanco, Astrid Xiomara
Tipo de recurso:
Doctoral thesis
Fecha de publicación:
2023
Institución:
Universidad de los Andes
Repositorio:
Séneca: repositorio Uniandes
Idioma:
eng
OAI Identifier:
oai:repositorio.uniandes.edu.co:1992/68869
Acceso en línea:
http://hdl.handle.net/1992/68869
Palabra clave:
Waterflooding
Optimization
Machine Learning
Deep Learning
Reduced-order models
Diffusivity Equation
Reservoir Engineering
Ingeniería
Rights
openAccess
License
Atribución 4.0 Internacional
Description
Summary:In the face of increasing global energy demand and the need for energy transitions, improved decision-making processes in the oil and gas industry are essential. Waterflooding is a successful method for enhancing oil recovery. Numerical reservoir simulation software are essential tools for evaluating the waterflooding process before its implementation in a reservoir. However, this software can be expensive, requires extensive information, and it is challenging to calibrate and optimize. To address this issue, we propose a decision-making framework that employs machine learning models and metaheuristics for near-optimal well control management, ultimately achieving maximum profit and effective oilfield management. Our research approach involves the dynamic reservoir evaluation and fluid production forecast using the diffusivity equation as a predictive numerical model. We reduce the computational time and resource consumption integrating a Proper Orthogonal Decomposition (POD) model to the diffusivity equation. With this model we reproduced the historical oil and water production rate based on the operational wells constraints. Due the high nonlinearity of the numerical predictive model and the challenge to incorporate it into the optimization algorithm we use machine learning models to predict oil and water production rates for each well under changing operational well constraints. The evaluated models are long short-term memory models, convolutional neural networks, and a combination of both. These models, along with the financial evaluation, are integrated into a non-linear optimization component. To solve this component, we use an Iterative Local Search, a metaheuristic that allows us to evaluate several scenarios and find a near-optimal solution. The decision variables are the bottom-hole pressure for each producer well and the water injection rate for each injector well. The objective function is to maximize the net present value. Overall, our proposed framework seamlessly combines the accuracy of the numerical predictive models, the computational efficiency of the reduced-order models, the advantages of neural networks, and the search power of metaheuristics. This provides an efficient strategy for waterflooding optimization over the mid and short-term in practice.