Method for improving accuracy of neural network forecasts based on probability mixture models and its implementation as a digital service

A method aimed at improving the forecasting accuracy is presented. It uses a combination of classical probabilistic-statistical models and neural networks. Moments of mathematical models are used as a nontrivial expansion of the feature space. The efficiency of the proposed approach is demonstrated by the analysis of several experimental data ensembles of the L-2M stellarator. Error decrease is especially noticeable when using the moments of the statistical models based on the increments of the initial observed data. To implement the methods of statistical analysis and the proposed machine learning algorithms, a digital service has been created. Its architecture and capabilities are also outlined.