Methods of applying neural network algorithms in forecasting of energy consumption level at systems of automated electricity distribution (In Ukrainian)

An analysis of modern methods of automated control of the level of electricity consumption at the level of households, industrial facilities, as well as critical infrastructure facilities, based on software algorithms and neural network architectures, was carried out. The exponential growth of the demand for electricity on a global scale and the globalization of the infrastructure of electricity networks indicate the urgency of the task of accurately forecasting the level of consumption, on the basis of which the optimal scheme for distributing electricity to consumers can be determined in real time. It is noted that when organizing a complex methodology for forecasting the level of electricity consumption and automated control of electricity distribution, it is necessary to establish statistical indicators that make it possible to estimate the volume of the input data array processed by the system, limitations on the calculation resource of the hardware and software complex of the platform, and requirements for the accuracy of machine analysis in accordance with financial risks and the probability of emergency situations. The high efficiency of the application of neural network infrastructures in the construction of systems of machine analysis, forecasting and automated control of the power grid infrastructure is shown. Such approaches to the organization of neural network architecture as a recurrent neural network, models based on long short-term memory and recurrent valve nodes, as well as time series models based on defined autoregressive integrated moving averages, according to which algorithms characterized by high accuracy of forecasting in real time under the conditions of minimal load on the computing resource. The importance of the preparation of the training selection and appropriate setting of neural network algorithms for the distribution of input data in accordance with the seasonal characteristics of electricity consumption is shown. The task of organizing, tuning and further optimizing the neural network algorithm was thus carried out according to the extrema of the objective functions, which were based on the statistical indicators of the prediction accuracy (mean absolute percentage error, root mean square error and mean absolute error) that were obtained from the results of the studies that were cited in open scientific publications.