Comparing the Impact of AI-Based versus Standard Load Profiles in ANN State Estimation Training in a Real Distribution Grid

Due to the increasing amount of renewable energy sources and new consumers in low voltage power grids, there is an increasing need for grid monitoring. For this purpose, there exist state estimation methods that can predict the unknown grid state. Historically, standard load profiles were used as input, but they may be outdated and not adequately represent small consumer dynamics. To improve the estimator, a novel artificial intelligence-based generator of load time series is used to generate small consumer load profiles. An estimation method is then trained on both standard- and novel load profiles to estimate the overall state of a realistic grid. The results of both runs are then compared with real grid measurements to determine the estimation error in each case. First results show that the overall estimation error is lower with novel synthetic load profiles. In line loading estimates for example, 52 % of upper quartiles were below a 5 % error with standard profiles, compared to 68 % of upper quartiles with novel profiles. Two topological errors in the grid model could also be identified.