Thermodynamic Consistency and Operational Logic Detection in Combined Cycle Power Plants through Explainable Gradient Boosting

Accurate forecasting of net electrical power output (PE) in Combined Cycle Power Plants (CCPP) is a requisite for grid stability and strategic bidding in deregulated energy markets. While Gradient Boosting Decision Tree (GBDT) architectures offer superior predictive precision compared to traditional physical simulations, their black-box opacity often hinders industrial adoption. This study presents a physics-consistency forecasting framework that combines a comparative stress-test of XGBoost, LightGBM, and CatBoost, with post-hoc SHapley Additive exPlanations (SHAP) analysis for thermodynamic validation. Utilizing a benchmark dataset of 9,568 operational hours, with an 80/20 train/test split and 5-fold cross-validation scheme, the optimized CatBoost architecture achieved the highest predictive fidelity, yielding a mean Coefficient of Determination (R2) of 0.969 and an RMSE of 2.97 MW during cross-validation. Furthermore, rigorous computational latency benchmarking (10,000 iterations) and jitter analysis confirmed CatBoost as the optimal engine for real-time Automatic Generation Control (AGC), demonstrating a sub-millisecond inference latency of 999 µs per sample. A pivotal contribution of this research is the unsupervised identification of a 24.6°C thermodynamic inflection point via bootstrap SHAP interaction analysis. This data-driven threshold is consistent with the onset of psychrometric performance degradation and the expected activation logic of auxiliary inlet air cooling systems, suggesting the model has internalized Brayton cycle thermodynamics without explicit physics programming. By mapping these data-driven inflection points to specific operational logic, this research demonstrates that a sensor-lean framework can achieve high maintenance resilience and function as a reliable data-driven operational surrogate for modern energy engineering.