(101c) Determination of Dispersion of CO2 Pipelines Using Machine Learning Approach

Carbon Capture, Utilization, and Storage (CCUS) is an emerging industry developed in response to climate change. As part of this effort, more pipelines are being constructed to transport CO₂ to sequestration sites. While CO₂ is non-toxic and non-flammable, it is an asphyxiant that can be lethal at high concentrations. If a CO₂ pipeline is accidentally ruptured, a significant amount of CO₂ could be released into the atmosphere, potentially posing a risk to human health, particularly in densely populated areas. Therefore, the assessment of the consequences of CO₂ from pipelines is essential to prepare accordingly emergency response for nearby communities to ensure safety. Given the large number of CO₂ pipelines and their varied conditions, a rapid, universally applicable prediction model is crucial for completing the tasks effectively. Computational Fluid Dynamics (CFD) is a powerful technique for simulating fluid behavior in complicated scenarios, allowing simulations considering complex phase transition and varying terrains. However, a disadvantage of CFD is its high computational resource demands, making it impractical to simulate every case. To address this limitation, the study integrates supervised machine learning to learn patterns and develop models for various terrains. The prediction models are based on the quantitative property-consequence relationship (QPCR) model, which connects the parameters of scenarios with corresponding consequence responses. The database used for training the models is derived from Ansys Fluent CFD simulations, based on the operating conditions, weather, and terrain of real-world high-pressure CO₂ pipelines. The machine learning models used to identify the best-performing model included multiple linear regression (MLR), support vector regression (SVR), k-nearest neighbors (KNN), random forest (RF), extreme gradient boosting regression (XGBoost), gradient boosting regression (GBR), and bootstrap aggregating (Bagging). Consequently, the R² score for all best models, as evaluated through 10-fold cross-validation, exceeded 0.93, indicating high prediction accuracy. The models enable CO₂ pipeline risk assessment to be integrated into both the planning stage and emergency response strategies.