Until late 2025, I had never thought much about the power grid. The prior eight years of my career were at Qualcomm, working on inspection systems for compact camera module production lines, catching defects and process deviations in real time, before they turn into six-figure scrap events.
I came across GFO-25-303 almost by accident. It's a CEC solicitation about improving net load forecasting by integrating behind-the-meter data, the problem being that grid operators can't see rooftop solar, home batteries, or EV chargers, and all that invisible generation is making net load increasingly hard to predict. I forwarded it to Sunki and said we should look at this, because the technical problem was one we already knew: performing root cause analysis on factors that aren't directly observable in the measurement you actually care about.
That needs some explaining. On a camera module production line, the hard part isn't catching defects you can see in the final image, it's tracing them back to a process step where you have no direct visibility. A tilt error on the finished module might come from adhesive dispense variation, or a thermal shift during curing, and you have to infer which one from downstream measurements. I had spent years building models that attributed visible outcomes to hidden upstream variables. Net load forecasting has the same shape. The grid operator sees aggregate demand at the substation, but buried inside that number is rooftop solar they can't meter, battery state-of-charge they can't observe, EV charging loads with no telemetry. The prediction part is not really the hard part; the attribution is.
We started looking at what the ISOs were actually running for their own day-ahead forecasts. Mostly regression models, some shallow neural nets. Fine on a mild Tuesday in April. But we pulled error logs from ERCOT during the 2023 heat dome and from ISO-NE during a January cold snap and the forecast errors were 3-4x worse than average. Those are the hours where the forecast actually matters, when ERCOT is paying $5,000/MWh instead of $30 and getting dispatch wrong is ruinously expensive. The ISO baselines were basically calibrated for the easy middle of the distribution and hoping the tails wouldn't be too bad.
Catching tail events was literally what we did for a living. Not average-case prediction, any regression model handles that, but catching the moments where conditions shift and the standard model breaks down. We'd just never applied it to energy before.
Sunki and I ran our first model on two years of CAISO load data using a setup I'd originally built for production line monitoring. It beat their day-ahead baseline by a wide margin, which honestly surprised us, we'd expected the energy domain to have quirks that would trip us up. We tried ERCOT because the market structure is completely different and figured it would be a good stress test. The numbers held. Then PJM, MISO, the rest. We spent about three weeks expecting it to stop working on the next grid, and it didn't. So we wrote it up.