A physics-based hybrid model for reliable gas turbine power prediction under extreme weather conditions

Accurate prediction of gas turbine maximum power output is critical for electricity market bidding and grid stability in renewable-integrated power systems. Physics-based models cannot adequately capture performance degradation caused by equipment aging, while data-driven models, particularly tree-based ensembles, exhibit poor extrapolation beyond training ranges, posing a critical limitation under increasingly extreme weather conditions. This study presents a practical hybrid residual-correction framework that combines a compact thermodynamic correction model with a ridge-regularized polynomial residual learner for data-scarce industrial deployment. The physics-based component provides a physically consistent baseline reflecting current equipment status, while the residual component learns systematic deviations not captured by the physics model. The framework was validated using 193 performance test samples collected over 4.5 years from an F-class gas turbine in South Korea. Under real out-of-distribution conditions, with the upper and lower 5% of observed ambient temperatures excluded from training, the hybrid framework achieves a test RMSE of 1.135 MW, outperforming a Linear-Ridge baseline by 1.71 × , a tuned multilayer perceptron (MLP) by 4.31 × , and tuned XGBoost by 5.83 × . Across three evaluation protocols (chronological split, out-of-distribution holdout, and five-fold rolling-origin cross-validation), the hybrid framework is the only model that ranks first under two protocols, and it never ranks below third. The MLP, although most accurate under rolling-origin cross-validation, degrades to 4.89 MW under real temperature extremes. A hyperparameter sensitivity grid places the default configuration within 0.4% of the best cell. In-distribution accuracy matches the Linear-Ridge baseline, whereas the framework's advantage lies in its robustness under real temperature extremes and its consistency across evaluation regimes.