A modified method for training hybrid networks based on distributed and parallel computing technologies has been developed, to improve the generalizability of the model, optimise data distribution and reduce dependence on the complexity of the subset, as well as to more efficiently collect results and correct parameters in the process of training neuro-fuzzy networks. It is shown that the use of distributed and parallel computing technology in the proposed method of training hybrid networks has made it possible to reduce time costs when using several processes simultaneously, including when using 12 processes, time costs are reduced by 3.28 times compared to the traditional (hybrid) algorithm and 1.07 times compared to the known algorithm based on distributed and parallel computing. It has also been experimentally confirmed that the distributed and parallel computing technology used in the proposed method of training hybrid networks reduces the time costs at each subsequent iteration. And with 1000 iterations they are 7.69 times compared to the traditional (hybrid) algorithm and 2.0 times compared to the known algorithm based on distributed and parallel computing. The proposed hybrid networks training method using distributed and parallel computing technology was tested in the task of regulating the fuel consumption of the TV3-117 turboshaft engine, the results of which showed an improvement in the quality metrics Accuracy, F1-measure, Recall up to 35.9 % compared to the traditional (hybrid) method for training hybrid networks, Multilayer Perceptron Classifier and other machine learning methods.