Development of Machine Learning Models for Multiple Leak Detection in Offshore Gas Pipelines under Single- and Multiphase Flow Conditions

Leak detection and localization in offshore gas pipelines is critical in preventing hazardous events and minimizing operational and economic losses. This study presents a structured machine learning (ML) framework capable of detecting and localizing single and multiple leaks under single- and multiphase flow conditions. Synthetic data sets were generated using OLGA software to simulate a wide range of realistic leak scenarios, including variations in leak size, location, pressure, and flow regime. Several ML models were developed, trained, and optimized, and their performance was systematically evaluated. For multiple leak detection, we developed and compared three stacked models: Voting Regressor 1 (VR1); Voting Regressor 2 (VR2); and stacking regressor (SR). The results show that, for the single leak detection model, the random forest regression model outperforms other models with an absolute relative error of less than 1.22% with noised data for single- and multiphase flow conditions. In addition, the SR model significantly outperforms the other models. Specifically, the SR model achieved R2 scores higher than 0.96 compared to less than 0.92 for VR1 and VR2, demonstrating its superior ability to capture diverse patterns and relationships in the data. These findings highlight the potential of advanced ML techniques to provide reliable and accurate leak detection solutions in complex pipeline systems.