Forecasting power flows in the electrical grid using graph neural networks:

Forecasting of power flows in the electrical grid enables grid operators to make advance grid calculations to recognize bottlenecks in good time and take countermeasures. Machine learning approaches using local forecasting models are not capable of considering dependencies between individual system components (such as transformers) in the grid, which is why the explicit knowledge of grid operators about system changes in the electrical grid is not included in the forecasting. Graph neural networks (GNNs) are capable of learning the interrelations of selected system components so that they can be used for forecasting the individual power time series.

The project is interesting to: Grid operators

• Forecasting vertical power flows at network nodes using GNNs and investigating the benefit compared to local models
• Data-based learning of interactions and dependencies between individual network nodes
• Publishing a paper

GNNs can be trained on graphs, making it possible to forecast a target parameter (in this case the vertical power flow) at the nodes of the graph. Dependencies between individual nodes are represented by edges and their edge weights. Various approaches are to be investigated to select suitable edges (edge weights). In particular, a data-based approach is planned to be examined in which the relationships between the nodes are learned in the form of edge weights. In order to map the individual time series characteristics of each network node in the model, a multi-task learning approach is to be added to the GNN.

As a global model, graph neural networks can provide benefits in various forecasting applications, with different node relationships being relevant depending on the use case. In forecasting vertical power flows on transformers, the idea of generating an abstract representation of the electrical grid using a graph stands to reason. Transformers are represented by nodes, whereas the edges between the nodes indicate whether two transformers are connected to each other via electrical conductors. This allows for interactions between transformers to be taken into account in the model. For example, maintenance work on a transformer can lead to one or more “neighboring” transformers taking over its task and showing a changed power flow as a result.

Whereas the project initially focuses on forecasting power flows on transformers, the graph can prospectively be enhanced by adding other grid components such as generation plants or electrical loads. The learning of edge weights would then make it possible to draw conclusions about the shares of various energy carriers in the vertical power flow or to recognize whether particular plants belong to a transformer. The great benefit compared to physical approaches is that complete mapping of the real electrical grid is not necessary, but individual plants or their aggregations can be added as nodes.

But the information flow between the nodes in GNNs can also bring advantages in applications where direct interactions between nodes play a less significant role. In forecasting the power generation of wind power plants or photovoltaic systems, for example, not only local weather information can be used, but also that of the adjacent plants, making phenomena such as storm fronts or cloud movements recognizable to the model.

• Preparing real and synthetic datasets
• Integrating a multi-task module into a GNN
• Evaluating the methods with defined edges and edge weights in the graph
• Enhancing the methods with learnable edges and edge weights
• Final evaluation of the methods

• TenneT TSO GmbH: Transmission grid operator TenneT supports the project by providing measurement data of vertical power flows.
  Contact: Dr. Matthias Gebhardt, https://www.tennet.eu

• University of Kassel (GAIN): The junior research group “Graphs in Artificial Intelligence and Neural Networks” (GAIN) at the University of Kassel has special expertise in the field of graph neural networks, providing advice and support in methodological questions as part of this project.
  Contact: Dr. rer. nat. Josephine Thomas, https://gain-group.de

Project period

01/12/2021 – 31/05/2022

Dominik Beinert, M.Sc.