Predicting Renewable Curtailment in Distribution Grids Using Neural Networks

Abstract

The growing integration of renewable energies into electricity grids leads to an increase of grid congestions. One countermeasure is the curtailment of renewable energies, which has the disadvantage of wasting energy. Forecasting congestion provides valuable information for grid operators to prepare and instruct countermeasures to reduce these energy losses. This paper presents a novel approach for congestion prediction in distribution grids (i.e. up to 110 kV) considering the n-1 security criterion. For this, our method considers node injections and power flow and combines three artificial neural network models. The analysis of study results shows that the implemented neural networks within the presented approach perform better than naive forecasts models. In the case of vertical power flow, the artificial neural networks also show better results than comparable parametric models: average values of the mean absolute errors relative to the parametric models range from 0.89 to 0.21. A high level of accuracy can be achieved for the neural network that predicts the loading of grid components with a F1 score of 0.92. Further, also with a F1 score of 0.92, this model shows higher accuracy for the distribution grid components than for those of the transmission grid, which achieve a F1 score of 0.84. The presented approaches show good potential to support grid operators in congestion management.