Speaker
Description
Multi-well placement optimization is a challenging task in the field development process of Geological CO2 Storage and Utilization (GCSU), as the objective function is multi-dimensional, discontinuous, and multi-modal. Despite advancements in gradient-based and gradient-free optimization methods over the past decade, the complexity of geological systems continues to hinder effective well placement optimization. Furthermore, the application of high-fidelity physics-based models for well placement optimization exacerbates the challenge due to the computationally expensive nature of running thousands of numerical simulations. To address these issues, we employ data-driven models (DDM) using various machine learning (ML) approaches to predict reservoir responses based on injector and producer well locations. This enables strategic multi-well placement to optimize hydrocarbon recovery and CO2 storage in partially depleted oil reservoirs while significantly reducing computational demands.
The Egg model serves as a benchmark case to validate the computational performance of DDM-based well placement optimization. Our formulation is focused on the CO2 water-altering-gas (WAG) operation, with well locations as input parameters and net cash flow (NCF) as the output after a specified operational period. Training datasets are generated using Quality Map (QM)-constrained random sampling (RS), with input features including well coordinates, permeabilities, porosities, initial saturations, pressures, time of flight (TOF), and well-to-well distances, while outputs capture cumulative oil, water, and CO2 production. We evaluate the strengths and limitations of various ML methods, including Multiple Perceptron (MLP), Extreme Gradient Boosting (XGBoost), and Deep Neural Networks (DNN), across different sizes of training and testing datasets. All these models achieved R² values exceeding 0.97, delivering fast and accurate predictions. Among them, MLP-based proxies stood out for their superior accuracy and computational efficiency, especially when applied to larger datasets.
Integrating the MLP model built upon 1,100 datapoints into a genetic algorithm (GA) allows effective optimization of the injector and producer locations in the studied heterogeneous, three-dimensional reservoir, with net cash flow (NCF) as the objective function. The MLP-GA outperforms traditional simulator-based GA optimization by improving computational efficiency threefold while achieving similar optimal well locations and NCF values, thereby supporting sustainable subsurface resource utilization.
| Country | Norway |
|---|---|
| Student Awards | I would like to submit this presentation into both awards |
| Acceptance of the Terms & Conditions | Click here to agree |
