THE COMED LETTER · ILLINOIS POWER & DATA CENTERS

The forecast finally caught up

A year ago this newsletter looked at PJM’s load forecasts for northern Illinois and concluded: “the flat prediction is probably wrong.” Twelve months of new data are in. The prediction was wrong — and PJM now agrees, loudly, in its own numbers.

The scoreboard

The chart below is the whole story. Each thin gray line is a PJM annual forecast “vintage” for the ComEd zone (published 2016–2024); the thick blue line is what northern Illinois plants actually generated, from PUDL’s cleaned EIA records. For most of a decade, every new vintage forecast less demand than the one before it, while actual generation quietly climbed past all of them.

ComEd-zone annual generation (GWh): PUDL actuals vs PJM forecast vintages. Actuals are the county-filtered approximation defined in the original analysis.

Then the correction arrived, in two installments. The 2025 vintage (orange) broke the decade-long pattern: for the first time since 2016, PJM forecast ComEd demand meaningfully above its previous vintage — about 104,800 GWh for 2028, versus the 91,400 the 2023 vintage had penciled in. The 2026 vintage (violet) went further: 133,800 GWh for 2030 — 47% above what the 2023 vintage projected for the same year. The “AI bump” that was missing from every forecast we reviewed in 2025 is now the dominant feature of the forecast.

Grading the September 2025 observations

  1. HELD “Past generation shows a steady, modest rise.” Still true, and less modest: ComEd-zone actuals reached ~109,400 GWh in 2025, up from ~105,000 in 2023.
  2. HELD “Hyperscaler growth is not an illusion; states and consumers are feeling the price increases.” PJM’s capacity auction for 2026/27 cleared at a record $329.17/MW-day — up roughly elevenfold from $28.92 two auctions earlier — with data centers driving most of the increase. Illinois’s consumer advocate now calls data-center rate reform urgent.
  3. VINDICATED “PJM forecasts do NOT show an expected AI bump and are relatively flat.” True of every vintage available to the 2025 analysis (2016–2023). See below for the plot twist about the vintages that weren’t available.
  4. CALLED IT “The flat prediction is probably wrong.” PJM has since rewritten its ComEd outlook upward twice, and its own 2026 report leads with significant long-term load growth.

Forecast accuracy, re-scored

MASE (Mean Absolute Scaled Error) compares each vintage against actuals over the months where both exist; below 1 beats a naïve forecast, above 1 loses to it. The deterioration the 2025 analysis flagged has continued: the 2023 vintage now scores 1.96, the worst on record. The young vintages score better so far — but with 17 and 5 overlapping months respectively, those grades are provisional.

MASE by vintage, recomputed with actuals through May 2026 using the original method. Lower is better.

The mix, one more year on

The generation-mix picture that anchored the original analysis extends another year without changing its character: nationally, gas dominant and growing, coal fading, renewables grinding upward; in Illinois, nuclear still towering over everything else. The last few months are provisional in EIA’s data and typically revise upward.

US monthly net generation (TWh)
Illinois monthly net generation (TWh)

The map

Illinois power plants (sized by cumulative generation, colored by primary fuel) and the data centers from the original survey. The geography still tells the policy story: nuclear and data centers cluster in ComEd territory up north; the big fossil fleet sits downstate.

Receipts

Fresh citations supporting the year-old call, all 2025–2026:

Methods, unchanged on purpose

Everything above was recomputed with the same transforms as the September 2025 analysis so the grades are comparable: PUDL’s cleaned EIA tables (now through May 2026), the same fuel-category mapping, the same 26-county approximation of the ComEd zone, the same MASE definition, and PJM’s published COMED forecast sheets (now eleven vintages, 2016–2026). The original medallion pipeline — Prefect, dbt, DuckDB, Grafana, and the full Databricks-free writeup — is documented in the repo, with the refresh script alongside it. The frozen September 2025 database is kept as-is; this page never overwrites its baseline.