![](https://digilib.itb.ac.id/assets/icon/pdf.gif)
2024 TA PP REFYDO MUHAMMAD FARHAN 1-ABSTRAK
Terbatas  Suharsiyah
» Gedung UPT Perpustakaan
Terbatas  Suharsiyah
» Gedung UPT Perpustakaan
The increasing demand for oil, driven by global population growth and industrialization, speeds up the depletion of oil resources and worsens environmental issues. CO?-WAG (Water-Alternating-Gas) injection not only reduces carbon emissions but also enhances oil recovery. However, optimizing the operational parameters for CO?-WAG injection to achieve desired outcomes remains a significant challenge.
This study investigates the application of machine learning models to optimize CO?-WAG injection processes. The aim is to develop predictive and optimization models using machine learning algorithms such as polynomial regression, neural networks, and Extreme Gradient Boosting (XGBoost). This is achieved by creating 1000 simulation models of CO?-WAG. Seventy percent of these models were used as training data, while the remaining 30% were used for validation of the optimization methods. The optimization process utilizes genetic algorithms and particle swarm optimization to find the best operational parameters. The final R² results indicate that the Neural Network and XGBoost models perform the best, with an R² of 0.99 in training and 0.97 in testing.
The developed predictive models of CO?-WAG are capable of quickly optimizing operational parameters for different reservoir conditions within the range of the training datasets. Sensitivity analysis reveals that the most influential features affecting Net Present Value (NPV), from high to low, are PVI (Pore Volume Injected), Injection Rate, and Injection Cycle. From the optimization, it is found that the best prediction is achieved with a Neural Network, deviating only 1% from the simulation results. XGBoost combined with PSO shows only a 7% deviation from the simulation results. Additionally, there is a 135% increase in NPV after optimization using XGBoost-PSO and a 117% increase in NPV using Neural Network-GA compared to the base case.