Planning LV grids by predicting residual loads of households via methods of machine learning

The use of decentralised energy resources (DER) in private households will change the load and feed-in behaviour in the future and thus influence the power flow at the low-voltage level. Therefore, grid planning, especially at a low-voltage level, must be adapted. Hence, new approaches for modelling the residual load of a household, with regard to different portfolios of DER, must be addressed. For this, an existing mathematical optimisation model to optimise both, a household's investment decision in DER and their operating times, is extended using machine learning (ML) approaches. Several methods of ML have been tested, however, the highest solution quality was achieved using neural networks. Following this, the prediction of the investment in DER by classification algorithms with convolutional neural networks shows the highest agreement with the deterministic approach. In the case of the prediction of DER operating times and residual loads based on them, recurrent neural networks with long short-term memory are used. Finally, this study indicates that the resulting predicted time series of residual load can be used for power flow calculations in the low-voltage grid. This allows grid calculations to be performed depending on existing or future political and market conditions.