An oil and gas pipeline liquid loading prediction method based on category-aware regression generative adversarial networks

Liquid loading in oil and gas pipelines drastically reduces transportation efficiency and threatens operational safety. Complex gas–liquid two-phase flow patterns make traditional mechanistic and regular machine learning models unable to meet engineering accuracy. This work presents a novel Category-Aware Regression Generative Adversarial Network (CAR-GAN) for liquid loading prediction. The proposed model employs a category-aware multi-branch adversarial structure, which eliminates feature interference across flow patterns and integrates flow-pattern identification and regression prediction into an end-to-end framework. The model consists of a flow-pattern sorter, multi-branch hidden layers, and dedicated discriminators and regressors to learn flow-specific features and capture cross-flow commonalities. Using 1200 balanced samples covering six flow patterns from public databases and OLGA simulations, the proposed model is trained and verified. Compared with five traditional neural networks and two mechanistic models, the proposed model achieves the best performance with MAE of 0.0270 and R2 of 0.943, and only one mispredicted non-loading case. Field validation in Changqing Oilfield confirms its strong robustness, offering reliable technical support for flow assurance in oil and gas pipelines.