Estimating MFN trainability for predicting turbine performance

Abstract The present paper examines the relationship between the mapping nonlinearity indicators ( distribution angle α and distribution gradient β) of training samples and the empirical trainability of MFN (multilayer feedforward neural networks) with the model problem being the mapping between turbine efficiencies and blade throat design. An empirical trainability measure is defined as a means of representing the degree of difficulty involved in training an MFN. The raw training samples are preprocessed using two and four different options for input and output components, respectively. These options result in mapping cases with different mapping nonlinearities and trainabilities. The results of a numerical experiment confirm that α and β can be correlated to the empirical trainability of MFN in the context of the model problem.