Conditional Tabular GConditional Tabular GAN-Augmented Ensemble Learning for Off-Design Performance Prediction of a Gas Turbine-ORC Combined Power Cycle Under Limited Operational Data

Abstract

Gas turbine–organic Rankine cycle (GT-ORC) combined power systems operate over a broad off-design range determined by the ambient state and the load demand, but in recently commissioned or prototype plants experimental operating data for model calibration and validation may be limited. Here, a framework is proposed that combines conditional tabular generative adversarial network (CTGAN) and synthetic minority oversampling technique (SMOTE) data augmentation with a stacking ensemble of five gradient-boosted and bagging-based models (Random Forest, XGBoost, LightGBM, CatBoost, Ridge-meta-learned stacking regressor) for combined cycle power output, thermal efficiency, and ORC net power predictions under scarce and noisy data conditions. A physics-based thermodynamic model is used to create the benchmark dataset from which a limited (N = 200–300) subset is sampled to represent the conditions typically available. SMOTE 3× augmentation is found to yield good results consistently across all models compared to CTGAN, with the best combinations reaching R 2 = 0.9998, 0.9997, and 0.9906 for CC_power_kW, eta_CC, and ORC_power_kW, respectively. SHAP analysis shows the ensemble models learn feature–target relationships that are physically meaningful. To our knowledge, this is the first demonstration of CTGAN-augmented ensemble learning for off-design performance prediction of GT-ORC systems, representing a transferable data-efficient machine learning solution for surrogate modeling of energy systems.