Half of ERCOT's hours missed by under 2.3%; the storm week missed by 18%

How big is a typical miss once you measure it against the size of the forecast? Each row is one hour of January and February 2021 on the Texas grid, with the forecast error divided by the forecast. A 2,000 MWh miss is a rounding error at 70,000 MWh and a real problem at 35,000. The percentage view shows that most hours were close, and that Winter Storm Uri was not a few bad hours but a bad week.

Interactive chart available when JavaScript loads.

Most hours missed by a few percent

The tall bars sit left of the median line. The median absolute percentage error was 2.26% of forecast. The 90th percentile of absolute percentage error was 14.23% of forecast. The bars beyond the second line are the long tail, and the worst hour missed by 39.5% of forecast.

Interactive chart available when JavaScript loads.

The tail is the storm week

Each point is one hour, placed by the size of its forecast. The storm-week hours sit high and to the right: the forecast expected record winter demand, and the grid served far less. During the storm week, the mean absolute percentage error was 17.5% of forecast. 70.2% of storm-week hours missed by more than 10%. 16.7% of hours missed by more than 10%.

Interactive chart available when JavaScript loads.

The week after the storm was still off

The cumulative curves separate three periods. Before February 14 the curve rises almost straight up; before the storm, the median absolute percentage error was 1.63% of forecast. The storm-week curve barely leaves the floor until 10%. The week after sits between them: after February 20, the median absolute percentage error was 7.73% of forecast, as the forecast kept expecting demand that had not returned.

Interactive chart available when JavaScript loads.

Takeaway

Measured against the forecast, the day-ahead model was within a few percent in a typical hour and within 14% in nine hours of ten. The storm week missed by almost a fifth of the forecast, hour after hour, and the forecast took another week to settle. Quote the median for the model and the storm week for the risk.

Limits

Percentage errors divide by the forecast, not by served demand; the storm-week numbers would be larger against actual. Served demand during the storm reflects rolling outages, not what Texans wanted to use. March contributes five hours.