Grid Load Sheds: htm-monitor on GridShedWatch

The problem, the data, the idea

The problem. A few times a decade a grid cannot make enough power, and operators have to cut electricity to customers on purpose. In Texas in February 2021 that left more than 4.5 million homes and businesses without power, some for days. The public got advance notice in only four of the eight events since 2020.
The data. Every US grid operator reports to the U.S. Energy Information Administration, every hour: how much power people used, how much the operator forecast the day before, and how much it generated. It is public and free (EIA-930).
The idea. htm-monitor learns what normal looks like for each grid, hour by hour, and keeps learning as it goes. When the numbers stop looking normal, it says so. No labels, no retraining, no weather data, nothing the operator does not already publish.
The pitch. Not a replacement for the operator, who sees far more. An automatic second opinion from public data, which does not share the operator's blind spots.

Two levels, like the weather service

GridShedWatch scores warnings the way the National Weather Service does: events warned in advance, lead time and false alarms, read together. Every method gets the same alert budget, chosen without looking at the events.

Watch

Something unusual is building

Early and less certain. Counts if it starts in the 7 days before the load shed.

About 12 a year per grid, roughly one a month.

Warning

This grid is in trouble now

Rare. Counts from 7 days before the load shed to 2 days after, since a warning during the event still helps.

2 a year per grid.

htm-monitor's only job in the benchmark is to say, every hour, how unusual each of five signals is. GridShedWatch's shared rules turn that into watches and warnings, and its scoring is the same for every method, so htm-monitor cannot be tuned toward the eight events. Full protocol: GridShedWatch scoring.

Next to the other methods

Events caught, of 8 (held-out: of the 5 events scored once after the rules were frozen). htm-monitor and the LSTM are averaged over 5 seeds. [a–b]: 5th–95th percentile with events resampled. Lead: median hours before onset among caught events. Operators: public notices in the 7 days before onset.

The honest read. htm-monitor is within about one event of the best method at both levels, and it beats chance at both, but it does not lead either: the LSTM catches a few more watches (with far less lead time), and the seasonal z-score catches more warnings. No method leads at both. What differs is the kind of emergency: the simple rules caught the Uri events days ahead, htm-monitor and the LSTM caught the Elliott events hours ahead, when the public got no warning at all. With eight events, differences of about one event are not meaningful.

Event by event

Each chart runs from 7 days before the load shed to 2 days after. The top row is what the operator announced. Each row below is one htm-monitor seed: where its watches and warnings started.

Public notice Notice to the industry Day only published Watch Warning Load shed begins

Limits, stated plainly

California's 2020 heat wave is mostly missed (a watch on 1 of 5 seeds). The heat was well forecast, so the forecast-error signals never broke. Only the operator's public Flex Alert warned it.
SPP is fragile. Watched days ahead on 2 of 5 seeds; which seeds catch it varies.
Hours of notice, not days, for most of Elliott. Duke and LG&E-KU got watches about 8 to 9 hours ahead; Tennessee Valley is the exception, about 4 days.
A replay, not live operation. Each hourly value is treated as available at its time stamp; EIA publishes with a short delay, so live alerts would come about an hour later.
Eight events, seven from two storms. This is a careful case study, not a detection rate. GridShedWatch adds each new load shed as it happens.