Learning Probability Distributions over Georeferenced Distribution Grid Models

When grid planners design updates for existing infrastructure in power grids, they frequently encounter a lack of trustworthy and readily-usable digital grid models. This is especially the case at the low-voltage (LV) level. While the location of secondary substations and end-consumers is often known, the topology is less certain and cannot be uniquely estimated. This work proposes a probabilistic framework to efficiently sample possible georeferenced grid topologies. A parametric probability distribution assigns an exact probability to each possible, georeferenced grid topology using characteristic features. The parameters of the probability distribution can be learned from known exemplary grid topologies. A Markov chain Monte Carlo (MCMC) algorithm is then designed to sample from the learned distribution with low computational complexity, thereby enabling efficient statistical inference. The described steps are demonstrated for the probabilistic modeling of a LV distribution grid in Schutterwald, Germany.