Analysis of configurations of LSTM networks for medium-term vector forecasting

The paper analyzes 36 configurations of LSTM (long short-term memory) architectures for forecasting with a duration up to 70 steps based on data whose size is 300–500 elements. For probabilistic approximation of observations, a model based on finite normal mixtures is used; therefore, the mathematical expectation, variance, skewness, and kurtosis of these mixtures are used as initial data for forecasting. The optimal configurations of neural networks were determined and the practical possibility of constructing high-quality medium-term forecasts with a limited training time was demonstrated. The results obtained are important for the development of a probabilistic-statistical approach to the description of the evolution of turbulent processes in a magnetically active high-temperature plasma.