A self-attention deep regression framework with generative data augmentation for foam-plunger optimization lift control in gas wells

Efficient deliquification is essential for maintaining shale gas well productivity under liquid-loading conditions. A self-attention LSTM–WGAN-based prediction method is proposed for early liquid-loading prediction and adaptive operating-mode selection of a plunger-foam composite deliquification system. The novelty is to establish an AI-driven adaptive foam-plunger control framework for gas well deliquification. Density-based spatial clustering method is employed to remove abnormal records, and the data after outlier removal are normalized and augmented using a Wasserstein adversarial generation model with gradient penalty. A self-attention-enhanced long short-term memory prediction framework is then established to forecast the oil–casing pressure difference and water–gas ratio, which is used to determine the operating mode of the composite system. Production data from six wells in the Changning Gas Field are used to develop and evaluate the proposed method. Results show that the proposed model accurately captures the dynamic evolution of both oil-casing pressure difference and water–gas ratio. Compared with baseline models, the mean absolute error is reduced by 12.50% for oil-casing pressure difference and 8.60% for water–gas ratio, which improves deliquification efficiency and enhances shale gas production sustainability.