An Open Day-Ahead Load Forecaster on 43 U.S. Balancing Authorities: Ahead of the Published Forecasts, Not of What They Could Be (Cardinal Grid Notes, No. 3)

Abstract

A day-ahead load forecaster built only from public data (hourly demand from Form EIA-930 and the NOAA HRRR, NBM and GFS weather forecasts available at 9 a.m. of the day before) was developed on eleven U.S. balancing authorities (BAs) and run unchanged, day by day, on the 43 BAs of its station map over 2023 to 2025. On the 38 BAs whose published day-ahead forecast is comparable with their demand, its mean daily absolute percentage error is 2.41 % against 3.16 % for the BAs' own forecasts, weighted by demand, and it is better in 26 of them; on the 27 it never saw during development, 3.18 % against 4.14 %. Two results qualify the margin. Most of it is a systematic hourly offset between what the published forecasts and the demand measure: the same forecasts divided by their own recent error at each hour reach 2.19 %. And on the days of the winter storms and the largest peaks the open forecaster under-forecasts the daily peak by more than the operators do. Web version and plain-language version: github.com/cardinalgrid/notes; cardinalgrid.com/notes. Code: ba-load-forecast 1.0.0.