Prediction of Liquid Accumulation in Gas Wells to Forecast the Critical Flowrate and the Loading Status of Individual Wells

Abstract

Liquid accumulation is a major problem in gas wells. The inability of gas to lift coproduced liquids to the surface imposes back pressure on the reservoir, limits the ultimate recovery and ultimately kills the well if improperly managed. Therefore, accurate prediction of its occurrence and reliable monitoring strategy is key to effectively handling liquid accumulation in gas wells. In this study, machine learning algorithms were used to develop regression and classification models to accurately predict the critical flowrate and the loading status of individual wells. The regression models used are the feed-forward neural network and a least squares support vector machine models while the decision trees model was used as the classification model to characterize the loading status of the wells investigated. These models were validated using actual published data and it was observed that the feed-forward neural network performed better in predicting the critical rate compared to the least squares support vector machine model with an R² value of 0.9833, and thus was adopted. The feed-forward neural network model was further compared with other critical rate models; and a consistent result with least percent error of 5.571% was also observed.  Form this study, it is obvious that the neural network model provide benefits and good prospects in investigating liquid loading phenomena in gas wells compared to empirical models that apply so many simplifying assumptions.