When a grid gets congested, the usual remedies cost money: redispatching generation, or curtailing wind and solar that the system was built to accept. There is a third option that costs almost nothing - changing the topology of the grid itself, by splitting busbars and reconfiguring substations. Switching is free. The reason it is underused is that finding the right combination of switches is computationally brutal.
Towards Efficient Multi-Objective Optimisation for Real-World Power Grid Topology Control, written with TenneT and presented at ACM e-Energy 2025, is our attempt at that problem on a real network rather than a toy one.
The size of the search
The section of the Dutch transmission grid used in this work has more than 1,000 nodes. Security analysis tests hundreds of N-1 and N-k contingencies. At every timestep the agent chooses from roughly one million valid switching states, and it has to do this for 24 hourly timesteps of a day-ahead plan while satisfying every operational constraint.
No operator has time to search that space by hand, which is why in practice a small number of known-good configurations get reused.
Learn, then plan
Our approach runs in two phases. First, a reinforcement learning agent is trained on the operator’s historical days against a reward that combines the competing objectives. Second, a planning phase searches switching budgets — how many switching actions the operator is willing to accept — and returns a set of plans rather than one.
That second part matters more than it sounds. The objectives genuinely conflict: minimising N-1 line loading, minimising the number of switching timestamps, minimising how many substations are reconfigured. There is no single correct weighting of those, and pretending otherwise is how optimisation tools lose the confidence of the people using them. The output is a Pareto set, and the operator picks the trade-off.
What it achieved
Evaluated on TenneT’s own historical data, the approach solved 31 of 31 in-distribution days, against 24 for the best expert strategy. On out-of-distribution days — periods the agent had not been trained on — it solved 137 of 181, against 123. Plans are produced in roughly four to seven minutes, which fits inside a day-ahead planning workflow rather than requiring one to be redesigned around it.
To be precise about what that is: these are results on real operator data in day-ahead planning, validated retrospectively. It is not a system running in live control, and no European transmission operator is running reinforcement learning in live control today.
Where it has been discussed
We presented this work in the poster session of the Global System Operations Summit hosted by 50Hertz in Berlin, where Anton Fuxjäger also joined the panel “Solutions for Topology Optimization — what is available today”.
Authors
Yassine El Manyari, Anton Fuxjäger, Stefan Zahlner, Joost van Dijk, Alberto Castagna, Davide Barbieri, Jan Viebahn, Marcel Wasserer.
The paper is available via this link.
