LEAK DETECTION BY DYNAMIC SIMULATION IN NATURAL GAS PIPELINES WITH HYDROGEN

Abstract

Integrating hydrogen into current natural gas transportation systems is gaining attraction as an energy source target. Although pipelines remain the best and most cost- effective means of gas transportation, the addition of hydrogen alters the flow dynamics and raises the need for efficient monitoring. Leakage is one of the most important hazards in pipeline systems, since it could affect safety, environment, and economy. This work aims to understand the flow behavior in a gas pipeline transporting a natural gas-hydrogen mixture without and with leaks. At this study, various leak scenarios are modeled using OLGA, a dynamic multiphase flow simulator. Different leak sizes, gas compositions, flow rates, pipeline diameters, and inclination angles are modeled through a sequence of simulations. The leak is situated at a specific point in the pipeline. Box- Behnken design is used to conduct a sensitivity analysis, which includes low, medium, and high values across the five aforementioned variables. The primary leak indicators are pressure, temperature, and mass flow rate. Additionally, synthetic noise is included in the pressure and temperature data to replicate realistic sensor readings. Matrix profiles are used to analyze time- series events for anomalies, marked as leaks. Then, a machine learning leak detection model is developed from simulation data using a random forest classification model. The machine learning model is developed once based on temperature and another time based on pressure as leak indicator. The results emphasize that noise highly affects the outcome of the leak detection system. The temperature-based model proves to be more sensitive to noise and measurement uncertainty than the pressure-based model. The temperature-based matrix profiles cannot identify any leaks, once noise is added to the data. The pressure-based model can still identify leaks with noise, but its prediction accuracy is reduced. Leak size, pipe diameter, and distance from the leak also influence the leak detection. Leak detection becomes easier through pressure data for larger leaks, smaller pipe diameters, higher flow rates, and through sensors closer to the leak location. This thesis highlights the necessity of selecting suitable indicators in leak detection. A machine-learning model is proposed based on random forest classifiers, utilizing the most common and cost-effective sensors (pressure and temperature) to enhance leak detection in natural gas pipelines without and with hydrogen. This model can be used as a valuable tool in monitoring transportation systems and reducing energy and financial losses associated with leaks.