The paper presents generalized information on possible methods for preliminary processing of initial data and the efficiency of these methods for training neural network models. Differentcombinations for processing the initial sample, as well as different types of activation functions and architecture of neural networkscan be selecteddepending on the simulation goals and the specifics of the subject area. The efficiency of preliminary data processing has been established with neural network models of energy characteristics of centrifugal compressors. Eliminating outliers in the experimental data and normalizing the initial data allowedto increase the accuracy of the model by 1,5 % compared with the model trained on non-normalized data. The importance of processing and preparing data is confirmed by findings of foreign studies using data preprocessing to create neural network models in different subject areas.