Spaces:

---

gridfm

/

genco-challenge

Like 21

Running

Open the app: it runs in your browser

Team: Carmine Delle Femine¹ ², Leire Garin Atxaga¹, Asier Diaz-Iglesias¹, Juan Pablo Maroto Herrera¹, Ane Miren Florez-Tapia¹, Marco Quartulli¹, Izaro Goienetxea Urkizu² (¹ Vicomtech, BRTA · ² UPV/EHU) · Code · Model · Dataset

Contact: cdellefemine@vicomtech.org

TL;DR

• A grid lab in a browser tab. Pick a real Texas heat-wave day, a hurricane or a heat dome. GENCO predicts the
  optimal dispatch of a 2,000-bus Texas grid model in about a second; physics checks it; gridfm-trustlayer's calibrated
  bands say how far to trust each line. No server, no licensed solver.
• Every answer is checked. A full AC power flow certifies each dispatch; when the check fails, the page solves the
  exact AC-OPF from GENCO's prediction. 98.5% of 200 test what-ifs are certified, in 1.8 s at the median, and so are
  all 41 real peak hours that have a dispatch.
• What it teaches. Two 115 kV lines limit this grid in any weather. Under a +8 °C dome over Houston or Austin the
  exact solver finds no dispatch. Harvey shifted 50 km southwest leaves no dispatch in 8 of 25 damage draws, shifted
  northeast in none of 333.
• Monte Carlo, live. GENCO's speed lets the page run studies in your browser: a heat wave in 34-106 s, or 20
  damage draws of a hurricane in 4 minutes, with the share left without a dispatch and its 90% interval.

Why

What can a grid foundation model do if it adapts quickly to a new grid and new weather, and runs on a laptop, even in
a browser? We tested GENCO on AC optimal power flow (AC-OPF), the cheapest dispatch within every voltage, line and
unit limit, to analyse heat waves and hurricanes in real time.

The data

• The grid: TAMU's Texas2k, with 2,000 buses, 544 units and 8 areas: a synthetic grid on Texas's geography, not
  ERCOT's network.
• The model: GENCO small, fine-tuned from its GOC 2000 checkpoint in 3.5 hours on one GPU, on 23,297 what-ifs
  with exact AC-OPF labels (gridfm-datakit): random heat domes, load, wind, sun, prices and outages, and the damage of 12 historical
  hurricanes.
• Heat domes: each acts through four relations: load, line ratings (IEEE 738), wind stall and thermal derating.
• Real heat waves: the 52 peak hours of four heat waves (2022-2024): ERA5 temperature, wind and sun, ERCOT's
  load against that summer's peak (EIA-930), Henry Hub gas. Load and wind are measured, so the heat acts only on
  ratings and plant capacity.
• Real hurricanes: Harvey (2017) and Beryl (2024) on NOAA's HURDAT2 tracks, never seen by GENCO: 600 damage
  draws each, with the lines and units the winds take out and a damaged zone keeping 40-100% of its load. The track
  shifts up to 50 km and the wind speed ×0.9-1.1, on random load, wind, sun and prices.

How it works

1. Predict. GENCO reads the what-if and predicts every voltage and dispatch in one pass.
2. Repair. GENCO's raw prediction does not quite balance power. We keep its set-points (each unit's output and
   voltage) and solve the AC power-flow equations within the units' reactive limits: the point now obeys the physics,
   though some limits may still be broken. Steps 1-2 take 0.5 s on 2 CPU cores, about 1 s in a browser: that is the map.
3. Certify. If a limit is broken, we linearise the power flow at that point, solve a small LP for the smallest
   change of set-points that brings every line and voltage within its limit, rerun the full AC power flow, and repeat
   if needed. Certified means that power flow meets every limit (ratings + 1%, voltages within 0.005 p.u.).
4. Exact step, if needed. If the check cannot certify a point, the page runs an exact interior-point AC-OPF
   (MIPS, in WebAssembly) from GENCO's prediction; its optimum is the answer if it passes the same check. Otherwise the
   page says "no dispatch found": a flag, not a proof.

Each line carries a 90% band of its loading at the optimum (gridfm-trustlayer, split-conformal): a line that
may be at its limit is drawn as such.

Is it right?

Certified means checked. On 1,484 unseen heat-dome what-ifs, the check certifies 97.0%, 0.27% above the optimal
cost (median); on 1,149 hurricane what-ifs, 92.1%, 0.73% above. We picked the model and its settings after seeing
these test sets; our pre-registered hurricane test failed. With the exact step, a 200-what-if benchmark goes from 92.5%
to 98.5% certified, and the median stays at 1.8 s. On the 52 real peak hours, all 41 with a dispatch are certified,
27 by the check (0.2-1.1% above the optimal cost) and 14 by the exact step (2-3.5 s each, at Ipopt's optimum); none of
the 11 where Ipopt finds no dispatch is.

GENCO is what makes the browser's exact solver usable. Started from GENCO's prediction, it certifies 161 of the
200; from its standard start, 103, in 9.6 s at the median. The best open-source exact solver we could tune (ExaModels +
MadNLP) is about as fast as the cascade, 2.0 s at the median, but needs Julia, so it cannot run in a
browser.

The trustlayer bands are calibrated, and say when they are not. On unseen heat-dome what-ifs they cover 90.5% of line
loadings; of the lines truly at their limit, the point estimate alone calls 19% "below", the bands 3.5%. On real days,
bands calibrated on 129 other hot days covered 84-87%, below the 90% target, so the app marks them as indicative.

What we learn about the grid

Two 115 kV lines limit this grid in any weather. At the exact optimum, one in the Rio Grande valley is at its limit
in 99.6% of the 1,484 heat-dome what-ifs and one near Burnet in 63%; of 3,220 lines and transformers only 63 ever reach
theirs, and these two carry 83% of the limits hit. Both are at their limit at all 41 real peak hours with a dispatch,
and in 99-100% and 28-29% of the hurricane draws.

A +8 °C dome breaks it around Houston and Austin. We centred a 300 km dome on 67 places over Texas. At +4 °C the
grid absorbs it everywhere, at 5.3% more cost (median). At +8 °C, Ipopt finds no feasible dispatch in 15 places and
cannot decide in 4; its "infeasible" is a local verdict, not a proof.

At the real summer peaks, the failures are the low-wind hours. In 11 of the 52 real peak hours Ipopt finds no
dispatch (8 infeasible, 3 undecided), all within 2.1% of that summer's peak load. Among the 22 hours that close to the
peak, these had wind at 7% of capacity against 16% for the others (median; exploratory). Each gets a dispatch back
with 0.4-0.9 GW more wind, or with 1.3-6.7 GW less load. ERCOT served those peaks; this says where a 2016-sized model
runs out of margin.

A hurricane breaks the grid depending on its track. With Harvey's track shifted 50 km southwest, toward Corpus
Christi, the exact AC-OPF finds no dispatch in 8 of 25 damage draws; shifted northeast, in none of 333. Beryl is the
mirror image. The draws that fail have about twice the lines out (21 against 12 for Harvey, 52 against 25 for Beryl);
load and wind do not separate them.

Checked live, in a browser. These findings come from offline exact solutions. In the app, a button reruns each
with GENCO on a second engine in the page, and every verdict matches: June 2023 (13 days, 34-45 s), August 2024
(13 days, 104-106 s), a +8 °C dome at 16 places (2 minutes), Harvey shifted 50 km southwest (20 damage draws,
4 minutes). That Monte Carlo finds no dispatch in 7 of 20, 35% (90% interval 20-53%), against 8 of 25 offline. On a
busier machine one rescue ran past the exact step's 15 s limit (19 of 20).

Open science

The app and its code are in our Space, the fine-tuned GENCO in
our model repo, the labelled what-ifs in our dataset.
GENCO is Apache-2.0, Texas2k CC BY 4.0, ERA5, EIA and HURDAT2 data public. Next: the same platform on an open Iberian network, for heat waves and
wildfires.

_{GENCO (arXiv 2608.09921) · gridfm-graphkit · gridfm-datakit · gridfm-trustlayer ·
TAMU Texas2k · ERA5 (Open-Meteo) · EIA-930 · NOAA HURDAT2 · IEEE Std 738}