Pipeline Leak Detection Combining Machine Learning, Data Assimilation Approaches and Pipeline Fluid Flow Physics Models

Abstract

With growing worldwide consensus about the impacts of climate change, the oil and gas industry faces unprecedented pressure to minimize its carbon footprint. The biggest source of carbon emissions in the industry is the so-called fugitive emissions, accounting for ~57% of the total oil and gas industry emissions, resulting from leaks in oil and gas pipelines and facilities. Fast, accurate and economic prediction of leaks in pipelines would significantly reduce fugitive emissions by reducing the time to respond to a leak. The proposed leak detection algorithm is a mixture of state-of-the-art machine learning and data assimilation techniques with well-known physical models and correlations of fluid flow in pipeline networks. The algorithm is tasked to continuously oversee pipeline operations by means of pressure and flow measurements. The proposed algorithm can probabilistically detect when and where a leak is taking place at the frequency of data collection (minutes/hours), thus minimizing the time to respond and the total fluid loss (fugitive emissions). The proposed algorithm utilizes a variant of the ensemble Kalman filter for probabilistic data assimilation together with an underlying network physics model. The model is augmented with meta-models and anomaly detection machine learning algorithms for real-time detection of leaks. The effectiveness of the proposed algorithm is demonstrated through a synthetic test case based on a realistic dataset.