Abstract
Early detection of a liquid loading has always been a challenging task. Typical existing best practices are the "cease to flow" problem occurs when a well is unable to remove the liquids produced from the wellbore. These liquids accumulate at downhole, forming a static column that creates back pressure against the formation pressure, leading to a decline in production until the well stops producing.
Primary technical challenges include the complexities of reservoir dynamics, variability in well and fluid properties, and environmental factors. Additionally, procuring labeled training datasets for machine learning models and interpreting these models specifically for well cease-to-flow or liquid loading issues are significant obstacles. Initial models used raw features such as allocated volumes for each production phase, real-time information, well test data (wellhead pressure, flowline pressure), and historical problematic events.
To ensure early detection of well cease-to-flow events, two distinct machine learning models were developed. These models assess the potential occurrence of such events using documented historical data. Events were categorized into three classes (healthy/normal, pre-event, and cease-to-flow conditions) based on reported flow-down code events, encompassing the entire history of these codes for accurate model labeling. The multiclass classification model was constructed to identify issues related to well cease-to-flow, with the goal of estimating the well cease-to-flow class. The primary aim of this model is to detect events preemptively, validated through blind testing on completions not included in the model's training set.
This solution has been successfully tested and implemented in a giant mature oil field. The chosen optimal model has a positive recall of 75%, meaning that out of 100 instances of well cease-to-flow events, 75 are accurately predicted. Conversely, the negative recall is 96%, indicating that out of 100 instances classified as "non-well cease-to-flow," 96 are correctly identified. This highlights the model's effectiveness in minimizing false alarms (4%) while successfully forecasting most problematic events.
Adopting a standardized approach significantly enhances profitability by optimizing intervention costs and maximizing incremental oil production (by reducing downtime and deferred production volumes). Improved opportunity selection optimizes workover schedules, reservoir monitoring, business plans, and intervention schedules, thereby enriching the knowledge base across the entire asset. Regular re-training of the machine learning models is crucial to accommodate the variability and evolution of field and reservoir conditions over time.
