Operational regime dependence of model performance and feature importance in machine learning-based gas turbine power prediction

Industrial gas turbines play a vital role in modern power generation, providing flexible and efficient power generation that supports grid stability as renewable energy penetration increases. While machine learning has been widely applied to gas turbine power prediction, existing studies typically evaluate models over a single, mixed operating regime and do not examine how feature importance and model rankings change once transient and steady-state operation are distinguished. This study addresses that gap using approximately two years of real industrial operational data from a continuously monitored gas turbine, comparing Linear Regression, Random Forest, XGBoost, and an Artificial Neural Network on both the original, mixed-regime dataset and a physically motivated, steady-state filtered subset (Fuel Flow Rate ≥ 65 KPPH, equivalent to 8.19 kg/s). On the original dataset, ANN and Random Forest achieved the best performance (RMSE = 3.94–3.97 MW), while Linear Regression performed substantially worse (RMSE = 46.7 MW) due to a pronounced nonlinear idle-to-operating transition. The ranking reverses when steady-state operation is isolated, with Linear Regression achieving the lowest error of all four models (RMSE = 1.74 MW, versus 2.14–3.69 MW for the nonlinear models). This regime-dependent reversal has not previously been reported in the gas turbine literature. SHAP-based feature importance further reveals that the four models show substantial disagreement regarding the dominant predictor under mixed conditions (individual model shares ranging from 19.1% to 75.1%, spanning turbine outlet temperature, fuel flow rate, and compressor exit pressure depending on architecture) but converge consistently on compressor exit pressure (45.2–82.7% of total SHAP importance across all four models) once steady-state operation is isolated, a convergence explained here through a quantitative variance-collapse mechanism rather than asserted qualitatively. A progressive feature elimination analysis further shows that, under steady-state conditions, a minimal one to three-feature model is not merely sufficient but measurably outperforms the full feature set (RMSE ≈ 2.6 MW versus 3.55 MW), an actionable finding for reduced-sensor deployments. These results demonstrate that both model selection and feature interpretation in gas turbine power prediction are fundamentally regime-dependent, providing a physically grounded framework for reliable, interpretable, and deployable predictive models in industrial gas turbine systems. The methods themselves are standard; the novelty lies in the paired original-versus-filtered comparison design, which reveals regime-dependent behavior invisible to the single-regime evaluations common in prior gas turbine studies.