This paper presents a process for automated start value generation for gas turbine performance simulations using Kriging metamodels. The metamodels are trained on a small number of pre-selected operating points in the flight envelope. Predictions of the trained metamodels are used as initialization parameters for subsequent performance simulations of arbitrary operating points in order to increase robustness and computational speed of the numerical process. Different approaches for the selection of the training points are evaluated. A comparison to the classical approach of table-based initialization is carried out to highlight the advantages and disadvantages of the new methodology.
Furthermore, the inclusion of supplementary operating points into the training sample of the metamodels is analyzed. Depending on certain criteria, such as the difference of prediction and simulation result, operating points are included into the sample and a retraining of the metamodels is performed. Simulations using the retrained models as guess value generators are compared to the previous Kriging approach. The advantages and disadvantages of the retraining approach are discussed.