The rising share of flexibilities combined with the recent development of machine learning techniques has the potential to create a sustainable and efficient low-voltage grid infrastructure. The new operational challenges, as the increasing uncertainty of the production of renewable energy sources, have to be met by effective measures. To prevent or mitigate disturbances such as overloading of the operating equipment or excessive voltage deviations, a predictive operation based on an autoregressive model using Graph Neural Networks is presented in this paper. The proposed model enables an optimized grid operation over a period of 1-5 hours by curtailment of generation and consumption at individual grid nodes as well as the adjustment of transformer tap changer positions.

A Graph Neural Network, which was trained with load and generation profiles for grid operations, serves as the basis of the model. It determines a grid operation for any possible combination of load and generation which can be obtained by current measurements or by forecasts for load and generation for each individual household in the grid. The load and generation forecast as well as the operating strategy determined by the graph neural network in the previous time step serve as input for the autoregressive model. With this information, the subsequent grid state can be calculated and an adjusted graph created. This graph is loaded into the Graph Neural Network, which determines the optimal operating strategy. The strategy is passed on to the next time step through the autoregressive approach. Thus, it determines an operating strategy over a series of consecutive time steps.

The autoregressive approach enables predictive grid operation over multiple time steps that influence each other sequentially. For example, it can be taken into account that the generation of photovoltaic systems is lower if a reduction was determined in the previous step. The model is tested on a pypsa model of a real existing grid with high shares of renewable generation and flexible loads. For the evaluation of the autoregressive graph neural network algorithm, a ground truth is generated for the relevant time period. This ground truth is determined by sequentially testing various operating configurations for each time step in a freely selectable order until a disturbance-free configuration is found. The optimal solutions obtained in this way serve as a reference. The results show a clear dependency of the algorithm on the training quality of the underlying Graph Neural Network. With a suitable selection of training and test data, further work aims to improve the Graph Neural Network, which will increase the quality of the autoregressive model. This straightforward and intuitive architecture can prevent faults in real time by regulating controllable generation and loads and prevent the overloading of lines and transformer stations in the grid at an early stage.