Energy efficiency modeling considered chemical process anomalies using contrastive learning-guided generative adversarial imputation network for operation-aligned data reconstruction

• A generative adversarial imputation network with contrastive learning is proposed • Contrastive learning drives imputations toward positives and away from negatives • GAIN uses identified pairs to reconstruct data for improving energy prediction • The proposed model can precisely predict energy efficiency with a test R2 of 0.9585 • Positive/negative pairs substantially improve generalization and imputation error. Effective industrial energy management in petrochemical operations relies on dependable measurement sensors; however, process data frequently exhibits systematic missingness due to sensor malfunctions and multirate sampling, which adversely affects the estimation of specific energy consumption. To address this problem, this study proposes a self-supervised contrastive learning that integrates with a generative adversarial imputation network for an LSTM energy efficiency predictor (LSTM-SSCL-GAIN). Positive/negative pairs pull imputations toward physically similar operating conditions and push them away from dissimilar regimes, guiding the generative adversarial imputation network to produce operating-consistent data based on the modified adversarial, reconstruction, and contrastive losses. The self-supervised contrastive learning integrates with a generative adversarial imputation network reconstructed data is then passed through the long short-term memory energy efficiency model to perform the prediction. The proposed model is verified by a large crude-glycerin refinery. According to the results, LSTM-SSCL-GAIN outperforms machine learning and deep learning models by achieving a testing r-squared of 0.9585, a root mean square error value of 216.36 kJ/kg, and a percentage error value of 1.04% in energy efficiency prediction. Robustness analysis indicates that the model with full contrastive loss maintains high accuracy in both reconstruction and prediction, even under missing rates of up to 60%. Moreover, the ablation study further shows that combining both positive and negative pairs delivers the highest performance and stronger generalization. Through operational adjustments, the proposed model indicates a 10.43% reduction in energy efficiency, equivalent to an annual emissions reduction of approximately 1,550 tons of carbon dioxide.