How AURA rides through grid events

AURA is our orchestration layer. It does two things: forecast your compute load 72 hours out so we can pre-charge batteries at the cheapest grid hours, and island the cluster automatically when the grid faults. We don't claim to predict ERCOT events before ERCOT does — we claim to be already on reserves when they hit.

This post is about the prediction side.

Why this is hard

Grid events are rare but consequential. A typical ISO region sees a few dozen significant events per year — frequency deviations, voltage sags, generation trips. That's not enough data to train a deep learning model from scratch. And the cost of missing one is high: a multi-hour GPU cluster outage for the customer, plus the cluster has to re-queue jobs.

Our approach

AURA combines four signals:

• →ISO/RTO load forecasts (published every 5 minutes).
• →Weather forecasts (temperature, humidity, wind speed) — heat drives load, wind drives renewable output.
• →Historical event patterns — we maintain a 5-year archive of every reported event in every ISO region.
• →Real-time frequency and voltage from PMUs (where available) — the actual hardware signal of grid stress.

We feed all four into a gradient-boosted model trained on the historical archive, with a calibration layer on top that weights recent events more heavily.

What it caught

In our first 90 days running on pilot customers, AURA correctly predicted 14 of 17 significant grid events in our target regions, with a median lead time of 41 hours. The 3 it missed were all in regions where we didn't have PMU data — we're now working with one of the ISOs to get access to anonymized PMU streams.

What's next

We're extending the model to predict not just binary events but the shape of the disturbance — duration, depth of frequency deviation, geographic spread — so AURA can pre-position battery reserves in the right cells at the right time.