Scaling Machine Learning Methods for Topology Control Congestion Management

Grid congestion in the Dutch transmission grid is already causing several large negative effects, such as blocking new developments and electrification. As trends in electrification and renewable energy production continue, these problems are likely to worsen. Traditional countermeasures, such as redispatching and/or curtailment, are costly and limited in applicability. Transmission system operators (TSOs) are therefore exploring alternatives.
One such alternative is topology control, which involves dynamically changing the network structure of the power grid, rerouting electricity and thereby resolving congestion. However, while promising, it is also hard: the number of possible reconfiguration actions is massive, and their effects are difficult to predict. Grid operators are rarely aware of which topological actions to apply. European TSOs are therefore starting projects to bring support tools for topology control into their control rooms. At TenneT, this tool is called GridOptions: it recommends topological actions to operators.
Currently, such tools use brute-force calculations. Such calculations are too slow, however, for wide applicability. Recently, research has focused on machine learning (ML) to address this shortcoming. Through learning complex patterns in grid data, ML models can learn to predict topological actions at a much higher speed.
While promising, further work is necessary to bring ML approaches into practice. In particular, real-world grids are much larger than synthetic experimental grids. Further research is hence required to scale ML methods to such larger grids. Similarly, grid operators have other operational objectives and requirements not accounted for in academic research.
In the proposed PD trajectory, we aim to bring machine learning models for grid topology control into large-scale application. Our primary focus is on scaling to larger grids. Additionally, we consider the organizational challenges of operationalizing such methods at a TSO and how operator requirements can best be incorporated into developed solutions. The research will be implemented at TenneT.