An Interpretable Predictive Model for Liquid Holdup of Hilly Terrain Oil-Gas Pipelines Based on Machine Learning and Shapley Additive Explanations

Abstract

During oil and gas field development, hilly terrain pipelines for mixed oil and gas transportation often face challenges from liquid loading, which significantly impacts flow assurance. However, the intricate nonlinear relationship between liquid loading and pipeline system operations complicates the solution of liquid holdup in practical pipelines by using traditional mechanistic models. Research combining machine learning (ML) for liquid loading prediction is limited, and most studies focus solely on validating the accuracy and efficiency of ML models without addressing their interpretability. To address this issue, an interpretable predictive model integrating ML and Shapley additive explanations (SHAP) is proposed. Operating conditions for an actual hilly terrain oil-gas pipeline were simulated using OLGA software to generate a comprehensive data set. Six ML models were compared using four evaluation metrics to assess their predictive performance for liquid holdup. eXtreme Gradient Boosting (XGBoost) demonstrated the best performance, with an R² of 0.986, MAE of 0.012, MAPE of 0.030, and RMSE of 0.016 on the test set. The XGBoost model was then visualized using SHAP to provide interpretability. Specifically, global explanations were given to ascertain the contribution of individual features to the average liquid holdup in low-lying areas of the pipeline. Local explanations were also conducted for individual pipeline operation data samples, visualizing the influence of each feature on the average liquid holdup within any specific sample alongside their cumulative effects. The interactions among the features were analyzed to derive trends in the interaction effects of different pipelines and operating conditions on the average liquid holdup over different ranges.