Abstract:In this paper, geometric deep learning techniques are applied to learn approximate models for power system estimation and calculation tasks. Nine different graph neural network architectures from literature are compared for this purpose. The underlying graph and known physical algebraic equations are taken into account during training, which allows to learn inductive, approximate algorithms for state- and powerflow estimation. The learned models are applied on randomly generated synthetic electrical medium voltage grids that are generated based on typical grid properties. It is shown, that the learned models are able to estimate the system states with high accuracy and that the trained neural networks can be applied to previously unseen grid topologies. Different sensor configurations and assumed random sensor noise and failures are taken into account. The trained neural networks are able to estimate the states with high accuracy despite of high sensor failure rates as well as noise that is added to the system.

Physics-informed geometric deep learning for inference tasks in power systems