QuicoptUse case · Transmission grid
Grid topology optimization
Wind and solar are being built out to meet rising electricity demand. Every day, Germany’s transmission grids have to move the volumes traded on the exchange without putting the network at risk.
On most days of the year, though, those grids are congested. Operators then have to throttle generators sitting in an unfavourable spot and start up others closer to the load. In Germany this is called redispatch.
Redispatch 2025
Redispatch and compensation for curtailed generators cost around 1.6 billion euros in 2025, according to the Bundesnetzagentur.
The sum of redispatch with conventional plants (1,176 million euros) and compensation for curtailed renewables (433 million euros), preliminary figures. Source: SMARD (Bundesnetzagentur), „Maßnahmenvolumen im Gesamtjahr stabil“, 30 March 2026
Switch instead of curtail
One option is to optimize the grid’s topology. Lines are switched in or out to steer the power flow and bring overload energy down.
Schematic illustration of the principle, not measured data.
Switching costs little next to redispatch: no fuel burned, and no compensation paid to curtailed generators. The CO2 balance improves at the same time, because less wind and solar has to be curtailed while conventional generation comes down.
The scale of the problem
A transmission grid of this size typically has 10,000 switches, and every day the operator has to pick the roughly five best switching actions out of them. That is a decision problem.
One hour after market close
Market close
15 minutes
produce and check proposals (IT systems)
45 minutes
manual validation and implementation (operator)
In the German market, topology optimization gets one hour a day after market close. The IT systems have 15 minutes of it to produce qualified proposals and check them. The operator keeps the remaining 45 minutes for manual validation and implementation.
What makes this hard
Too many combinations
The optimizer has to pick about 5 actions out of 10,000 switches. Trying all 10¹⁸ possible combinations is out of the question. This is a hard problem.
N-1 security is mandatory
Every topology has to be N-1 secure: if a single component fails, the rest of the grid must still run safely. So every topology proposal, every proposed switch setting, has to pass the time-consuming N-1 check first.
DC first, then AC
To speed this up, the first pass uses the simplified DC representation. Every topology proposal is then re-validated in AC and discarded if it does not hold up.
All within 15 minutes
The N-1 check, the optimization proposals and the AC validation all cost time. The challenge is to produce as many good and workable topology proposals as possible inside those 15 minutes, to bring overload energy down and avoid redispatch costs.
The Quicopt solution
Quicopt’s OR solver algorithm computes optimal switch settings and returns good, workable proposals in a short time. Because the N-1 cost function in DC is not available as a simple algebraic formula, MILP solvers and gradient methods run into their limits, or cannot be applied directly at all.
N-1 cost function: the black box can be connected as an oracle function. You never have to disclose your data.
Fewer N-1 calls
The N-1 check is called comparatively rarely, which saves time in an already tight 15-minute window.
A short solution list
Quicopt returns a short list of solutions. That cuts the time the AC validation needs inside the 15-minute window.
Runs on standard CPUs
The algorithms run on standard CPUs; GPUs are optional. They integrate into existing IT environments without much effort.
Developed at FZJ
The method was developed at Forschungszentrum Jülich by solver and power-grid experts, and computed on a public transmission grid. We are in touch with transmission system operators.
Quicopt can therefore deliver more qualified topology proposals in the time available. That helps bring overload energy down at low cost, and with it the need for redispatch the following day.
Benchmark
Numbers can be checked, claims cannot. We ran Quicopt against the production prototype of a German transmission system operator on two public PEGASE test grids. Both sides solve the same task on the same data, and every topology one side reports is re-scored by the other.
| Grid | Action budget | Δ overload energy | Quicopt | Comparison system |
|---|---|---|---|---|
| case1354pegasebase | 4 splits + 3 disconnections | 0.0000 | 1.4 min | 20.4 min |
| case1354pegasewide | 4 splits + 3 disconnections | −165.49 | 2.1 min | 20.3 min |
| case1354pegasewide | 4 splits + 3 disconnections | 0.0000 | 2.1 min | 60.3 min |
| case1354pegasebase | 4 splits | 0.0000 | 1.0 min | 20.2 min |
| case1354pegasebase | 6 splits | 0.0000 | 1.0 min | 20.4 min |
| case1354pegasewide | 6 splits | −20.99 | 1.2 min | 20.2 min |
| case1354pegasewide | 6 splits | −8.99 | 1.2 min | 60.2 min |
| case9241pegase | 4 splits | −510.70 | 4.3 min | 21.6 min |
| case9241pegase | 4 splits | −126.81 | 4.3 min | 80.9 min |
| case9241pegase | 6 splits | −1,663.40 | 4.3 min | 21.7 min |
| case9241pegase | 6 splits | −474.59 | 4.3 min | 81.5 min |
Δ overload energy in MW.
The column shows the difference Quicopt minus comparison system. A negative value means Quicopt reaches less overload energy, and less overload energy is better. 0.0000 means both sides reach exactly the same value. Where the same task appears twice, the comparison system was given three to four times the compute in the second row.
What the action budget means
The action budget says how many switching actions the operator is allowed to execute. Splitting a busbar and disconnecting a line are capped separately. A larger budget sounds like more freedom, but it is the markedly harder problem, because the number of combinations jumps with every additional action. That is exactly where the margin grows: on case9241pegase it is 510.70 MW at four splits and 1,663.40 MW at six.
One run, several budgets
Quicopt answers several action budgets from a single computation. The four case9241pegase rows come out of one run of 4.3 minutes. The comparison system computes each budget on its own.
And the time window?
For the 15 minutes the IT systems are left with, this is the difference that decides. Quicopt fits inside it at 1.0 to 4.3 minutes. The comparison system needs 20 to 82 minutes across its four runs.
The hardware
The hardware is not the same on both sides, and that belongs in the picture: the comparison system runs on one A100 GPU, Quicopt on 64 CPU processes of a 128-vCPU machine. Standard CPUs instead of GPU accelerators.
The test grids
- case1354pegase (base): 1,354 buses, 1,991 lines, 6,515 possible switching actions, 1,741 contingency scenarios.
- case1354pegase (wide): the same grid, but 43,527 possible switching actions.
- case9241pegase: 9,241 buses, 16,049 lines, 98,423 possible switching actions, 14,384 contingency scenarios.
Quicopt computes heuristically. Every result was validated with the other side’s own code.