Development and validation of a neural network for state estimation in the distribution grid based on μPMU data

The increased expansion of decentralised renewable energy resources (DER), such as photovoltaic (PV) systems, and private charging points for electric vehicles (EV) with high charging capacities of up to 22 kW pose new challenges for distribution grid operators in Germany. In Germany, both DER and EVs are installed in the electrical grid primarily at the low-voltage level. In order to ensure grid stability in the future, energy generation by DER and charging of EVs must be coordinated locally. At the low-voltage level, however, there is mainly no measurement technology installed which is required for the implementation of a grid-compatible control of the charge EVs and feed-in of DER. It is not economical to equip every grid node with measurement technology. The paper presents an artificial neural network (ANN) which is trained by means of the voltage data of a software image of a real low-voltage grid and which subsequently estimates the voltage of all nodes of the grid section on the basis of a few measuring points in the real low-voltage grid. The ANN has been developed and validated in software. An average accuracy of the actual node voltage to the estimated node voltage of the ANN of 4.29 V mean absolute error, 1.23 % mean absolute percentage error and 4.49 V root mean square error has been achieved.