Early detection of closed-loop slugging patterns in offshore oil wells with unsupervised learning approaches

In offshore wells, severe slugs are frequent problems that can limit oil production. It has already been proven that active pressure control can mitigate this effect, but defining the setpoint is still a manual task that requires constant human intervention. This study proposes a novel approach utilizing machine learning to aid the search for optimal production levels while preventing severe slugging. Two unsupervised machine learning methods, namely Self-Organizing Maps (SOM) and Generative Topographic Mapping (GTM), were evaluated for the early detection of slugging patterns in offshore oil wells. This study utilizes simulated FOWM model data to construct the necessary database. Additionally, real-world well data underwent SOM and GTM analysis, providing valuable insights from practical contexts. Both SOM and GTM showed promising results. However, GTM outperformed SOM in terms of mapping orientation and prediction scores. In addition, the GTM was easier to optimize in terms of hyperparameters for map tuning.