Abstract

Liquid loading refers to undesirable accumulation of liquids (oil and water) in the wellbore. This typically occurs in late-life wells as the reservoir pressure declines. However, in unconventional reservoirs, liquid loading is also found to occur in fairly productive oil wells upon shut-in, leading to extended downtime and production loss. Unfortunately, it is hard to know a priori which wells would liquid load when shut in. In this work, we provide mechanistic insights as well as develop a machine learning model to quantify the risk of liquid loading in naturally flowing wells so that preventive actions can be taken to minimize the downtime.

This work first leveraged transient simulations (OLGA) to probe the mechanism of shut-in induced liquid loading in unconventional wells with long laterals. Based on the acquired insights, various machine learning models, including tree-based algorithms (random forest and gradient boosting), feed-forward neural networks, and recurrent neural networks (simple RNN, LSTM, and GRU), were implemented and trained using historic data of a number of lateral wells that liquid loaded. The best performing model was selected to predict for naturally flowing wells how likely they would liquid load if there were shut-in events. A standard machine learning operations (MLOps) infrastructure was established to automatically deploy predictions and maintain model performance in production.

Mechanistic insights were gained for shut-in induced liquid loading. Key findings include: 1) ups and downs along the well path in the lateral wellbore can lead to local gas entrapment by the liquid, which can obstruct flow restoration upon well re-opening after shut-ins; 2) multiple choking/opening cycles, often attempted in operations as an effort to restore well production, can worsen liquid loading. These findings suggested the importance of having features that characterize well lateral geometry in the subsequent machine learning modeling. In the prediction of liquid loading, recurrent neural networks were found to outperform other machine learning models due to advantages in handling time series data. Accuracy of model predictions was validated by ground truth data.

This work showcases an example of using both physics-based and data-driven modeling to maximize value from optimizing production and operations. For the first time, the mechanism of this troubling issue was unveiled for ExxonMobil Operations in unconventional reservoirs. More importantly, large potential for downtime reduction is enabled by easy application of the workflow to a large number of wells across unconventional resources that suffer from the same issue. This work may also promote the optimization of drilling, completion, and artificial lift system design for unconventional wells with long laterals. Lastly, the success of this end-to-end data-driven workflow will open the door to tackling other challenging problems in unconventional production and operations.

This content is only available via PDF.
You can access this article if you purchase or spend a download.