Notes from engineering and research.
Grid demand forecasting, deep learning architectures, and what operating across U.S. power markets has taught us.
When to publish nothing
Three ways a forecast dashboard can mislead you, and the operational rules we wrote so the live Gramm dashboard fails toward silence instead of toward a confident wrong number.
What March 2026 taught us about forecast anomalies
A four-day mid-March heat event spiked day-ahead load forecast errors on every major U.S. grid by 2-3x. How we detected it, root-caused it, and shipped a response that cut CAISO March MAPE by 1.07pp.
Benchmarking across seven U.S. grids: what we learned
Each grid has different characteristics, solar-heavy CAISO, isolated ERCOT, wind-variable SPP. One model architecture, trained per-region. Here is what we found expanding from California to all seven.
The architecture search that found something better
We benchmarked every modern deep learning architecture against seven U.S. grids. The winner was not what we expected. And by the time we submitted the paper, we had already found something more performant.
Why grid demand forecasts fail during extreme weather
Conventional forecasting methods break down during heat waves and cold snaps. The load-temperature relationship becomes nonlinear at extremes, and ISO baselines are calibrated for normal conditions.
We had two papers and no product
Early papers suggested deep learning could beat ISO baselines, but later live tests forced a stricter evidence standard. This is the story of turning research into an API without hiding the gap.
Why we invented a new accuracy metric before building the model
Two models with identical MAPE can have wildly different operational consequences. We borrowed from statistical process control to define metrics that distinguish nuisance errors from dangerous ones.
How a semiconductor team ended up forecasting power grids
Gramm's team came from semiconductor manufacturing, attribution of visible outcomes to invisible upstream variables. When we read the CEC's GFO-25-303 solicitation, we recognized the problem immediately.