Building a Hybrid Computational Fluid Dynamics - Machine Learning Framework for Real-Time Pipeline Leak Detection in Oil and Gas Systems

Abstract Pipeline leak detection is a great challenge in Nigeria's oil and gas sector. Fugitive emissions from leaks account for approximately 57% of the Oil and Gas Industry's total carbon footprint. While Traditional inspection methods which are hardware based sensors and pressure monitoring software are constrained by high installation cost and false alarm rates there is a need for smarter and scalable solutions. This study proposes a Hybrid Computational Fluid Dynamics-Machine Learning (CFD-ML) Framework to achieve a cost-effective, real-time monitoring solution to Pipeline leak detection by utilizing transient CFD simulations of methane flow in Ansys Fluent to Machine learning Algorithm. The dataset was generated from computational fluid dynamics simulations using Ansys fluent. It consists of approximately 5000 rows of pipeline operating conditions under leak and no-leak scenarios. Sensor features comprising pressure, velocity, and turbulence intensity were extracted, and a scenario-based splitting was adopted to evaluate the model performance on unseen conditions. A Random Forest model was employed as a baseline model for leak detection, while leak localization was formulated as a classification and regression problem. The detection model achieved near-perfect performance on unseen scenarios. In contrast, the classification-based localization model failed to generalise to unseen leak positions, and regression-based localization produced predictions that were confined within the range of the observed training locations. These results indicate that localization performance is primarily constrained by sparse spatial sampling in the data rather than the model complexity. This suggests that improved localization requires denser coverage of leak positions.