METHOD AND ALGORITHM FOR EXTRACTING FEATURES FROM DIGITAL SIGNALS BASED ON NEURAL NETWORKS TRANSFORMER

The construction of a robust model and an assessment of its accuracy in problems of forecastingelectrical loads with additive consumption profiles are considered. A study was conducted on the influenceof neural network parameters (data packet size; number of neural network layers; neuron activation functions; optimizers) on the error in predicting power consumption. Graphs comparing the profiles of actualand projected consumption and the deviation of the forecast for electricity consumption above the averagevalue for the period under review are presented. Optimal parameters of the predictive neural networkmodel have been selected in manual mode. The result of the study of the varieties of genetic algorithmsrevealed the optimal hybrid algorithm for learning a neural network model based on the rapid convergenceof the solution. A Python-based algorithm for selecting network hyperparameters based on powerconsumption data with different patterns of electricity consumption has been tested. The conducted trainingand testing of the genetic algorithm confirmed the possibility of obtaining forecasts of greater accuracyand the possibility of automating the selection of optimal hyperparameters. In the tasks of forecastingpower consumption using a neural network model, regardless of the method of creating the structure, theoptimal metric has been selected. It is revealed that for consumers with additive profiles of electricity consumption,it is advisable to use the robust Huber loss function, at the same time, for consumers with aunique or regular profile of electricity consumption, the use of a sliding window increases the error, unlikeadditive consumers. It is shown that the use of a genetic algorithm significantly increases the accuracyof forecasting due to the individual selection of optimal parameters for a specific consumer. A blockdiagram of an intelligent device for predicting energy consumption modes has been developed. A decisionmakingassistance system has been introduced that allows for the implementation of planned proactivemanagement based on data taken from the electricity meter and obtained as a result of the neural networkforecasting model. The decision–making assistance system calculates the deviation of the projected powerconsumption values from the actual ones and, as a result, issues recommendations to the dispatcher of thedistribution power grids. Based on data from the decision-making assistance system, the distribution gridoperator can make a decision on ordering the required amount of electricity, gets the opportunity to monitorpossible spikes and decreases in consumer electricity consumption, abnormal equipment operation,and additionally monitor the adequacy of the neural network model