How does a new framework screen voltage control policies before they hit the real grid?

The gate that stops new controllers

A distribution grid operator who wants to try a learning-based voltage controller faces a hard gate. The controller must prove it will not push voltages outside allowed bands before it ever touches the physical network. Simulators help but they miss disturbances. They miss modeling errors. They miss device interactions that appear only on the real system. Historical measurements capture reality but only under the old control policy. A new policy can steer the grid into operating regions that have never been observed. These two gaps make a direct safety case impossible.

What DR-CSS actually does

Distributionally Robust Conformal Safety Screening (DR-CSS) treats the simulator as a nominal predictor. For each new operating scenario the simulator outputs a voltage trajectory for every bus in the grid. DR-CSS then wraps a conformal safety interval around that trajectory. The interval width comes from historical simulation-to-reality errors recorded while the old policy was running. Because the new policy changes the closed-loop dynamics, the interval is enlarged again to cover the shift induced by the new policy and its interactions with the remaining controllers. The framework is policy-agnostic. It does not need to know the internal structure of the new controller. It only needs the scenarios it will face.

The interval grows in two steps

First the conformal layer uses past simulation-to-reality errors. Then a distributionally robust enlargement accounts for closed-loop changes.

Why the second enlargement matters

When a new policy goes live it changes how the whole grid behaves. Other controllers react differently. Loads respond differently. The simulator trained on old data cannot predict these shifts. The robust enlargement bounds the worst-case shift within a Wasserstein ball around the historical error distribution. This ball radius is a tuning parameter. Larger radius means fewer false passes but more false alarms.

First framework of its kind

To the best of the authors' knowledge DR-CSS is the first framework in power systems to combine historical data from an existing control policy with an imperfect simulator for pre-deployment safety screening of a new policy. The paper was submitted by Sarra Bouchkati, Petros Ellinas, Adriana Geisler, Steffen Kortmann, Johanna Vorwerk, Spyros Chatzivasiliadis, and Andreas Ulbig. Bouchkati, Ellinas, and Geisler contributed equally.

Benchmarks show zero misses on unsafe cases

Experiments run on the IEEE 33-bus and IEEE 141-bus test systems. The target policies were learning-based voltage controllers. DR-CSS flagged every unsafe test scenario. No unsafe case slipped through. The raw method also produced warnings on some safe scenarios. To cut those false alarms the authors adapt the safety intervals to different operating conditions and introduce the new policy in stages with recalibration after each stage. The staged rollout raises the informational value of the screening and gives operators a clearer go-or-hold signal.

Open questions

The preprint does not say how many historical error samples are needed for a reliable interval. It does not quantify the computational cost of the conformal construction on a utility-scale model. It does not describe how the method behaves when the simulator itself is biased in a systematic way rather than just noisy. Those questions remain open for the next round of work.