Day-ahead demand forecast accuracy during peak-demand hours. Eight balancing authorities, September 2024 to September 2026.
Code, data, and the Power BI version: github.com/josealemanm/united-states-energy-grid-stress-analysis
Region codes. CISO: California ISO. ERCO: ERCOT (Texas). ISNE: ISO New England. MISO: Midcontinent ISO. NYIS: New York ISO. PJM: PJM Interconnection. SOCO: Southern Company. SWPP: Southwest Power Pool.
MAPE is the average forecast error as a percentage of demand. A stress hour is an hour in the top 5% of a region's demand within the same season. The stress penalty is MAPE during stress hours minus MAPE during normal hours.
This page shows the week around one hour: the extreme-demand hour (top 1% of demand) with the largest forecast error among hours that pass all data checks. It is Southwest Power Pool at 08:00 on January 26, 2026, local time. Demand was 45,510 MW and the day-ahead forecast was 57,910 MW, an error of 27.2%. In that hour net exports were 517 MW.
The pages above define a stress hour by how high demand is. This page also defines it by how fast net load is rising. The two definitions select mostly different hours. PJM and Southern Company are more accurate during high-demand hours but less accurate during fast increases in net load.
Question: when a balancing authority is in the top 5% of its own demand, is its day-ahead forecast error higher than in its other hours, and by how much?
Stress hour: an hour in the top 5% of demand for a balancing authority within the same season. Each region is compared only with itself, so regions of different sizes can be compared. Measuring within each season keeps the result from becoming a comparison of summer with winter. The top 5% gives about 438 hours per region per year.
Data checks: the table of largest errors on page 2 and the hour chosen on page 3 use only hours that pass every check below. All other charts use every hour.
Data handling: hours missing demand, forecast, generation, or interchange are excluded, not estimated. Hours with zero or negative demand are removed, and forecasts of exactly zero are treated as missing. Hours that fail the balance check are kept and reported, not corrected.
Limitations: MISO and SWPP cover more than one time zone but are each assigned one. The data is EIA's raw series, not its adjusted series, so reporting errors remain visible. The cost per MWh is an assumption from a published source, not a measured value. Demand depends heavily on weather, and 24 months is a short period.
Source: U.S. Energy Information Administration, Form EIA-930, API v2. Data pulled 2026-09-02.