To address the challenges posed by the complex and dynamic factors affecting the production and wellbore liquid loading in plunger gas-lift wells, a data-driven prediction model for liquid drainage and gas production in these water-breakthrough wells was developed. A dynamic simulation model for plunger gas-lift wells was established using the transient multiphase flow simulator, generating simulation models for various combinations of reservoir, wellbore, and production parameters. The Spearman rank correlation coefficient method was used to analyze the relationship and degree of association between reservoir formation factors, wellbore parameters, wellhead dynamics, startup/shutdown procedures and gas production rate, liquid production rate, and wellbore liquid loading. A convolutional neural network(CNN) was used for model training to create predictive models for gas production, liquid production, and liquid loading in plunger gas-lift wells. The model was deployed and validated in the gas well clusters at the Shen 11 station of Changqing Oilfield. The research and field application indicate that the OLGA-based simulation model for plunger gas-lift wells can effectively simulate transient gas production, liquid production, and liquid loading. The CNN-based proxy model demonstrated high prediction accuracy and is highly interpretable, serving as a technical foundation for optimizing the production and liquid drainage procedure of plunger gas-lift wells.