Study Of Surrogate Prediction Model And Application In Turbo-machinery Design

As the traditional stochastic optimization is time-consuming and also has more calculation, this paper studied a surrogate model methods with little calculation and acceptable accuracy based on the statistical theory for engineering design. In the course of turbo-machinery blade design, the surrogate model can also accelerate the design process and reduce the design costs than the traditional computational fluid dynamics analysis precision. Although there are universal commercial software (ISIGHT,etc) to provide surrogate model features, but for specific design problem(turbo-machinery engineering, etc), the open, customizable, modularized efficient agent technology is essential for further optimization of the overall design). This paper studied three kinds of surrogate model technologies, the main work can be summarized as follows:(1) Firstly, this paper studied the surrogate modeling mechanism, the experimental design strategy of optimal Latin Hypercube Sampling, and the model evaluation criterion of Kriging, RBF-NN, and S VMR model. However, when comparing with performance of model, the different dimension of the output made the results of RMSE different, so we could not judge the predictive ability of the model based on RMSE under different test conditions. This paper proposed the criterion indicators of relative root mean square error (RRMSE), and applied it to the comparative analysis for different surrogate prediction models.(2) We divided different test problems into four types according to the extent of nonlinear and dimensions of the design variables:high-level high-dimensional model, high-level low-dimensional model, high-dimensional low-level model, low-level low-dimensional, and divided the initial sample into sparse sample size, small-scale, large-scale sample according to the different size of them. Then we built108groups predictive surrogate models for12sets of test functions under different categories to predict and compare the results, we found that the predictive ability of different surrogate models presented differently in different types of predictions, but for the prediction in small-scale samples of high-level high-dimensional and high-level low-dimensional, RBF-NN model perform better than SVMR and Kriging model in prediction accuracy and robustness.(3) Based on above, we took hydraulic efficiency and head of mixed-flow pump as predicted target, keeping other design parameters and working conditions unchanged, establishing surrogate model with Kriging, RBF-NN and SVMR to predict the pump hydraulic performance in mixed-flow pump vanes design process. Then we took10groups of testing sample to test the prediction accuracy and robustness of the model, and compared RBF-NN model with the other two surrogate methods, we found that the RBF-NN surrogate prediction model is more accurately for the design of Turbo-machinery blades.Theoretically, we expected to give a surrogate computational model appropriate for complex design problems of the blade, and then predict the design goals’trends to the design variables through the model; In application, we expected to provide key technical supports for the design of complex turbo-machinery blades.