Liquid loading is a major factor limiting the production of gas wells. Liquid loading starts when the gas production rate does not provide sufficient transport energy to lift the liquid out of the well. The downhole liquid accumulation imposes an additional hydrostatic backpressure on the formation and decreases the production capacity of the well. This paper presents a data-driven model to predict the minimum gas rate required to lift the liquids out of the well.
A dataset consisting of 394 horizontal plunger-lift gas wells was used to train the data-driven model. A total of 32,292 points of liquid loading onset were found in the dataset and the corresponding gas flow rate was labelled as the target. Principal Component Analysis and Mutual Information Regression were performed to choose twelve most influential features on the critical gas flow rate (target). A model group consisting of three different machine learning algorithms and ten different data preprocessing steps was trained on the dataset. The most accurate model in the group was chosen to be the critical gas flow rate model.
Results showed that the data-driven critical gas flow rate model predicts the critical flow rate with R2 scores of 0.97 and 0.90 for train and test datasets, respectively. Also, the model results were compared to a few well-known empirical correlations in the literature (Turner, Coleman, Belfroid, Wang, Nagoo, and Shekhar equations). Results confirmed that the data-driven model has the highest agreement with the Belfroid model in comparison with the other equations. Next, an unsupervised learning method was used to modify the Belfroid equation. The modified correlation was proposed as a simple engineering tool to calculate the critical flow rate with higher accuracy in comparison with other correlations studied. It was observed that the Belfroid model underestimates the critical flow rate by around 32% and 5.2% for wells with tubing pressures lower than 300 Psig and higher than 300 Psig, respectively.
Most of the current critical flow rate models are based on two theories to explain the onset of liquid loading, droplet reversal and film reversal. Yet, the liquid loading mechanism is unknown due to the complexity of dynamic interactions between the reservoir and wellbore. Besides, most of the current models have their own limiting assumptions. This study provides a model that avoids unlikely theories or unrealistic assumptions and utilizes one of the biggest datasets of wells.