Application of decision tree algorithm for accurate prediction of liquid loading in gas-condensate wells

Liquid loading occurs when gas fails to lift co-produced condensates to the surface, causing backpressure, reduced production, and possibly resulting in well shutdown. This occurs when gas velocity falls below the critical level required to carry liquids, leading to their accumulation in the wellbore. Accumulation can occur in both vertical and horizontal wells, reducing efficiency, particularly in wet or retrograde gas wells. Accurate prediction and monitoring are crucial but often challenging due to the complexities of multiphase flow and estimating bottom-hole pressure. This study assesses the effectiveness of the decision tree algorithm for predicting the loading status of these wells, aiming to improve predictive accuracy and operational decision-making. Two decision tree models were developed using wellhead pressure and gas production rate as input features. The first model, with a maximum tree depth of 3, was designed to prevent overfitting by limiting the complexity of the decision tree. This constraint helped maintain model simplicity while still achieving an accuracy of 80%. The depth limitation ensured that the tree did not grow excessively, which can sometimes lead to overfitting, and instead focuses on capturing the most significant data patterns with a limited number of decision nodes. The second model, with constraints on node splits and leaf samples but no depth limitation, reached an accuracy of 78%. The results revealed that gas production rate is a more influential factor than wellhead pressure in determining well loading status, with the second model indicating that wellhead pressure becomes less relevant when the gas rate exceeds 75,365.2 m3/day. Both models performed well overall but showed potential for improvement. Future work should focus on enhancing model accuracy through advanced techniques such as ensemble methods and by increasing the dataset size through the inclusion of additional well data. Despite the limitations of a relatively small sample size, the findings underscore the potential of decision tree models in optimizing well productivity and operational efficiency in gas-condensate reservoir management.