Model education in the POD space for loss prediction in LPT cascades

The present work deals with post-processing by means of Proper Orthogonal Decomposition (POD) of data obtained from experimental investigations on Low-Pressure Turbine (LPT) cascades. The aim is to provide a strategy useful to coarsening design space grids and hence, reduce the data acquisition time during experimental investigations or high-fidelity simulations. The experimental work has been conducted in a wind tunnel installed at the Laboratory of Aerodynamics and Turbomachinery of the University of Genova. The experimental procedure consists in surveying the flow upstream and downstream of the cascade, aiming to evaluate the total pressure drop through the cascade, and hence the overall losses. Specifically, effects on losses of design parameters such as freestream turbulence intensity, vane solidity, incoming wake parameters (velocity defect, turbulent content and reduced frequency), flow Reynolds number and incidence flow angle have been evaluated. Then, POD has been applied to data constituted by total pressure loss coefficient distributions in the pitchwise direction under different conditions in order to: • highlight the loss trend vs the design space parameter variation • identify the minimum number of requested conditions in the design space, i.e., the number of independent conditions useful for the education of a predictive model These objectives have been achieved due to the POD capability in providing compact subspaces able to reproduce the same loss trend observed in the physical space. Moreover, in the POD subspace it has been possible to define a smooth kernel with the number of independent conditions in order to estimate the loss distribution of the remaining conditions. Hence, a model has been educated by means of Least Squares Method (LSM) on the POD design space coefficients choosing the predictors as low order polynomials of the independent variables. Results will show good agreement between measured and modelled loss distribution varying each design space parameter, showing that the model is able to predict loss distribution within the uncertainty measurement. Furthermore, keeping the same strategy for the education of the kernel on another dataset characterizing the loss distribution of two cascades with different geometrical parameters, loss distribution and magnitude are still well captured. Thus, analysis show that the rank of the problem is much lower than the tested conditions, and consequently a reduced number of tests are really necessary. This technique could be useful to reduce the number of scale-resolving simulations or detailed experiments without significant loss of accuracy.