Journal of Engineering Research and Reports

Natural gas is composed mostly of methane, the simplest hydrocarbon molecule, with only one carbon atom. But most gas at the wellhead contains other hydrocarbon molecules known as Natural Gas Liquids (NGL). Heavier gaseous hydrocarbons such as propane (C₃H₈), normal butane (n-C₄H₁₀), isobutane (i- C₄H₁₀) and pentanes, may also be processed in gas plants and exported as Liquified Natural Gas (LNG). During operational services in gas plant from inlet to outlet piping, gas leaks tend to occur undetected at some points in the facility. Apart from loss of gas resources, leaks and venting at natural gas processing plants release other pollutants besides methane (e.g., benzene, hexane, hydrogen sulfide) that can threaten air quality and public health. Hence, the need for early detection of gas leaks by using appropriate Machine Learning (ML) models. Insight from existing general flow equations was used to develop a new modelling approach for Machine Learning, in a test case: Gas Plant JK – 52. Input gas pressure data is calibrated and evaluated for consistency in real-time. The data is then corrected for lag-time and used to compute tolerance. Indicated time of alarm is checked against events such as residual gas, supply, pumping, etc. Where alarm is eventless, leak is suspected and eventually confirmed, suggesting that action should be taken to mitigate against the leakage. Following the input of a split training dataset, different types of regressions were used for the machine learning before automating the system for real-time evaluation and detection. Linear regression provided a 39% test accuracy, which was considered too low. This led to the use of random forest regression, which provided a 95% test accuracy and was considered excellent. It is hoped that with continuing data acquisition in gas plants employing this algorithm, further modelling will become more predictive as machine learns from experience.