The chapter is devoted to the choice of the optimal neural network architecture (NN) and the method of its training, providing forecasting with the least error. A multifactorial model of power consumption based on a multilayer NN has been synthesized and tested. Within the framework of the conducted research, an NN model was built describing the architecture of a cyber-physical system (CPS) for forecasting electricity consumption. It has been established that for each consumer, due to significant differences in the nature of energy consumption, it is necessary to experimentally select network parameters in order to achieve a minimum prediction error. It is shown that with atypical power consumption, i.e., not repeated over time periods (hour, day, week, etc.), artificial intelligence and deep machine learning methods are an effective tool for solving poorly formalized or non-formalized tasks. The developed model has acceptable accuracy (MSE deviation up to 15%). To increase the accuracy of the forecast, it is necessary to regularly refine the model and adjust it to the actual situation, consider new additive factors that affect the power consumption curve.