ASSESSMENT OF THE INFLUENCE OF NEURAL NETWORK HYPERPARAMETERS ON THE ACCURACY OF FORECASTING ENERGY CONSUMPTION

Abstract

The work is devoted to the problem of improving the accuracy of short-term forecasting of electricity
consumption using deep machine learning tools. The influence of the specified neural network NN
hyperparameters on the error in predicting power consumption, such as: data packet size – Bs; number of
NN layers – j; neuron activation functions – Fa; optimizers – O. The optimal hyperparameters of the NN
model for predicting electrical consumption (EC) for consumers of additive and cyclic types have been
determined. The analysis of the impact of the batch size on the accuracy of the forecast showed an increase
in the effectiveness of NN training with the growth of the input data package. The analysis of the
influence of the number of layers showed that with an increase in the number of layers of the NN, the
learning time decreases and its predictions become more accurate. A study of various optimizers for
learning speed has shown that the best results are demonstrated by the optimizers “Adam” and
“RMSProp". It is shown that the choice of the activation function determines how quickly the NN will be
trained and how accurate its forecasts will be. The use of various regularization methods allows NS to
achieve better results in practice, improving their generalization ability and increasing the accuracy of
predictions. It is shown that in order to achieve the minimum forecasting error, it is necessary to individually
configure the network parameters for each consumer, taking into account significant differences in
the nature of energy consumption. The training and testing of the created network with selected parameters
was carried out on a training and test sample containing data on electricity consumption for 2 years
(17520 hours). The analysis of input data on power consumption showed that the optimal parameters of the predictive neural network model in manual mode are: package size 250 (selected empirically), 5 layers,
activation function “ReLU", optimizer “Adam". Various ways of selecting hyperparameters (manually
and by means of genetic algorithm (GA)) are considered.