Prediction-focused machine learning for performance adaptation of aero gas turbines through steady-state and transient simulation

Abstract

Performance adaptation is a method that enhances the accuracy of aero gas turbine performance models through the use of measurement data, and is valuable for applications such as control, operational optimization, and diagnostics. In traditional methods, adaptation factors are derived using physical models and numerical analyses, but these approaches are computationally intensive and time-consuming. This study introduces a performance adaptation method based on prediction-focused machine learning for both steady-state and transient operation, an approach that offers significant efficiency and accuracy improvements over conventional techniques. Using a feedforward neural network, component adaptation factors are determined and used to refine the engine performance in the steady state, while correction factors are used to adjust the predictions under transient conditions, thereby enabling real-time adaptability. This integrated method minimizes prediction errors across the most critical performance parameters. For example, the average absolute errors are reduced to 0.14 % for the total temperature at the high-pressure turbine exit, 0.38 % for the fan pressure ratio, 0.15 % for the low-pressure shaft speed, and 1.11 % for the engine net thrust. These results demonstrate not only the precision of the proposed approach but also its practical advantages in terms of reducing computational costs and improving adaptability. This study addresses the limitations of traditional performance adaptation techniques, and a scalable and efficient framework is presented that significantly enhances the accuracy of aero gas turbine performance models, thus paving the way for more reliable and responsive engine modeling in both research and industrial applications.