Bridging Synthetic Training and Real-World Application: Applied Simulation-based Neural State Estimation

Abstract:
Most neural state estimation models in the literature are exclusively trained on a fully observable synthetic data set. Thereby they do not address the question of how to transfer this approach to a real unobservable grid, where the distribution system state estimation (DSSE) is supposed to be applied. This paper shows how to model a synthetic reproduction of a real medium voltage (MV) distribution grid. We show that we can use this as a basis for training a model on synthetic data and validate its’ performance using actual measurements of the MV grid. We propose a stratified date sampling method for training, validation and test set separation and a method for assessing real-world performance given a subset of measurements. Furthermore, we propose using dropout layers for estimating the uncertainty of the DSSE prediction.