Yu.V. Koltsov, E.V. Boboshko

The paper observes an approach for artificial neural network training with data, obtained from system modeling. This approach allows significantly broadening the area of artificial neural networks application with problems, which have not enough training data. It can be effective to apply the described approach for neural network training in cases, where observation for the wanted parameters of the real system is either difficult, or wasn-t preformed systematically. The main problem here is an appropriate modeling of the investigated system, by means of those variables, which can be observed well. Then, using the constructed model it is possible to get values of the parameters, required for training set formation. The process of modeling, wanted model-s parameters calculation and neural network training can be considered as preliminary preparation of data for the mass operation. As an example of the proposed approach, the problem of electrical energy losses calculation and forecasting in power distribution networks of 0.4-20 kV voltage is observed. Low observability and insufficient information about network operation regimes is common for these classes of networks. A numerical experiment was conducted in order to prove the effectiveness of neural network, trained on model data, application for losses calculations in a particular feeder of 6-10 kV voltage. The neural network approach found out to be more precise, than Average Loads method, which is normative for the concerned class of electrical networks.