Artificial neural networks to predict the node voltages in a low-voltage grid

Abstract:
In order to be able to balance the feed-in from distributed energy resources (DER) as well as the consumption by e.g. electric vehicles (EV) locally in the low-voltage (LV) grid, an intelligent control system is necessary. Such a control system requires forecasts of the state of the grid. This paper presents a forecast model which predicts the node voltages in a LV grid with only a few measuring points for different forecast durations. At these measuring points micro-Phasor-Measurement-Units (µPMUs) are installed. The forecast model consists of artificial neural networks (ANNs) trained with measurement data from a simulated LV grid. Load profiles and real generation profiles over three months are used as a basis for the simulation. The forecast model is validated with one month of simulation data. In the validation, the forecast model predicts the node voltages with an mean absolute error (MAE) from 0.897 V to 0.939 V for forecast durations between 6 and 24 hours. A ten-minute updated forecast reduces the MAE. Between the different forecast durations, there is no clear indication of the optimal duration. However, better results are obtained by updating. In unknown scenarios, which the ANNs has not learned, the forecast model does not provide sufficient predictions.