In-Context Whitening lets power flow surrogates adapt 21-34x faster
No gradient steps needed: one CPU core can reroute an ML grid model in milliseconds.
Machine‑learned surrogates for AC power flow (ACPF) excel on fixed networks but lose one to two orders of magnitude accuracy when a line outage changes the topology—an “operator shift” that alters the admittance matrix. Existing fixes require per‑topology gradient steps and fresh data, which is expensive at scale. Ayushi Jolotia and Parikshit Pareek propose In‑Context Whitening (ICW), a gradient‑free, weight‑free, architecture‑agnostic method that adapts by re‑estimating the output space’s whitening statistics from just a few hundred solved cases on the new topology. They prove ZCA whitening is the unique affine transform that preserves coordinate‑wise semantics, so two moments are sufficient for efficient correction.
On IEEE 30‑, 118‑, and 300‑bus systems under single (N‑1) and double (N‑2) contingencies, ICW reduces overall error by 6× to 28× over frozen surrogates (up to 54× per quantity under N‑2) and cuts worst‑bus power‑balance mismatch by 30×. It matches or beats gradient‑based adaptation in accuracy while adapting 21× to 34× faster. Crucially, the adaptation cost parallelises on commodity CPU cores rather than requiring one GPU per contingency, making it practical for grid operators who must quickly reroute power after outages.
- ICW achieves 6×–28× error reduction over frozen surrogates on IEEE 30‑ to 300‑bus systems under N‑1 and N‑2 contingencies.
- Adaptation is gradient‑free and 21×–34× faster than per‑topology gradient methods, running on CPU cores instead of GPUs.
- Proves ZCA whitening is the optimal affine transform that preserves physical output semantics with only two moments.
Why It Matters
Grid operators can now adapt ML power flow models instantly after outages, improving reliability without expensive retraining.