A Generative Adversarial Imputation and Hybrid XGBoost Approach for Refinery Corrosion Prediction

ABSTRACT

Refining units face severe multiphase flow corrosion, yet existing early‐warning systems suffer from missing monitoring data and poor model generalization. We propose a prediction framework combining a Generative Adversarial Imputation Network (GAIN) and an Improved Particle Swarm Optimization‐tuned Extreme Gradient Boosting model (IPSO‐XGB‐DROP). First, GAIN reconstructs missing data through physically constrained adversarial learning, preserving nonlinear physical coupling and ensuring physical consistency at up to 50% missing‐data rates. Next, IPSO‐XGB‐DROP introduces feature dropout into the optimized XGBoost to deliver precise predictions, jointly tuning hyperparameters and dropout rate to reduce overfitting. Validated on real industrial data, the model achieves R ²  = 0.944 ± 0.018, RMSE = 0.0115 ± 0.0012, MAE = 0.0095 ± 0.0008, demonstrating the integration of data‐driven and physics‐informed methods for predictive maintenance in high‐risk scenarios.