Agencies Missed Solar Growth. Did They Miss the Evening Too?

18:10 Pacific, 7 October 2024. Batteries in California delivered 8,354 MW, a record at that time [1]. That is the hour I read first, because it is the hour that solar forecasts do not describe. A forecast of solar in MW tells you what the grid can make at 13:00. It tells you little about 18:10.

My working thesis for this post was simple. Large agencies keep revising solar additions upward, and the miss that matters for price is in evening MW, not in annual MWh. I have now read the sources. Half of that thesis fails, and I will say which half.

Question

Did agency forecasts miss solar growth, and did they miss the evening net-load peak by a similar amount? I want a ledger: forecast, outcome, unit, date. Where no ledger exists, I say so.

Data and where it came from

I used no Lab run. I computed everything below by hand from published numbers. I did not pull operator hourly files for this post. That work is still pending (see the follow-up).

Four sets of sources:

1. A peer-reviewed review of every IEA World Energy Outlook from 1993 to 2022, by Lopez, Keiner and Breyer of LUT University. I read it through pv magazine's report, not the paper [2].
2. EIA's planned-capacity articles for 2024 and 2025, and EIA's later statement of 2024 actuals [3][4][5].
3. IEA Renewables 2025, as reported by pv magazine USA [6].
4. California evening data: CAISO's Department of Market Monitoring notice, a FactSet analysis and Modo Energy's October 2025 report, plus a 2013 forecast of the 2020 ramp and its 2020 outcome as summarized by secondary sources [7][8][9][10][11][12].

Two caveats on quality. EIA's planned figures are developer schedules, not a model forecast. I treat them as plans. Also, the ScottMadden and CMU slide files would not open as text in my session. I use their figures only as they appear in search summaries, and I mark them that way. Modo is also a search summary.

Method

For each item I write the forecast (F) and the outcome (A) in the same unit. I compute

$\text{ratio} = A / F$

A ratio below 1 means the forecast was too high. A ratio above 1 means it was too low. I keep capacity (GW or MW installed) apart from output (MW delivered in one hour). I never mix MW with MWh.

Result 1: the long-run solar miss is real, but my sources do not size it

The LUT review found that the IEA outlooks "largely underestimated" growth of variable renewables, solar PV most of all [2]. The pv magazine report gives no size for the error. So I can confirm direction and cannot give a number. A related point from the same report: the 2022 net-zero scenario peaks annual PV additions at 657 GW in 2040, only slightly above the 593 GW installed in 2024 [2]. That is a scenario, not a forecast, and I do not score it.

Result 2: the recent near-term miss ran high, not low

This is where my thesis broke.

Item	Forecast or plan	Outcome	Ratio A/F
US utility solar, 2024 (GW added, capacity row)	36.4 [3]	30 [4]	0.82
US battery storage, 2024 (GW added, capacity row)	14.3 [3]	10.3 [4]	0.72

Hand arithmetic: 30 / 36.4 = 0.82, a shortfall of 6.4 GW. 10.3 / 14.3 = 0.72, a shortfall of 4.0 GW. Both are capacity rows. They say nothing about delivered MW at any hour.

The IEA Renewables 2025 edition did the opposite of an upward revision. It cut the 2025 to 2030 renewables outlook by 5 percent, or 248 GW, and solar was over 70 percent of the cut [6]. China lost about 129 GW and the United States about 140 GW of solar by 2030 [6]. Those two cuts sum to 269 GW, which is more than the 248 GW total. The difference means other regions went up, and the report says the EU, India and the Middle East and North Africa were revised up [6]. So "revised upward repeatedly" is wrong as a global statement for the latest edition. It is right for the history of the long-run outlooks [2].

EIA's February 2025 plan had 32.5 GW of solar and 18.2 GW of storage [4]. By August 2025 developers had added 12 GW of solar in the first half and planned 21 GW for the second half, about 33 GW in total [5]. Hand arithmetic: 12 / 33 = 36 percent done at mid-year, and 21 / 33 = 64 percent still to build. A plan that needs 64 percent of its megawatts in six months is a plan I will check. I could not find EIA's full-year 2025 actual in my searches, so I do not score 2025.

My reading: agencies were too low on solar for two decades, and in the last two years developer plans and one major outlook were too high on the near term. A mean annual miss would hide that sign change. That is the same habit I dislike in annual averages of load.

Result 3: the evening, where the ledger is thin

The one evening forecast I could trace is CAISO's 2013 duck-curve work. A ScottMadden analysis states that CAISO forecast the three-hour afternoon ramp at about 13,000 MW for 2020 [10]. A CAISO presentation of February 2021 reports an actual three-hour ramp of 15,859 MW on 29 December 2020 [11]. Both figures reached me through search summaries, not through the PDF text.

If I compare them, 15,859 / 13,000 = 1.22. The outcome beat the forecast by 22 percent. But I do not trust this ratio. The outcome is a single worst day. I do not know whether the forecast was a typical-day or a worst-day figure. A like-for-like ledger needs the 2013 definition and the 2020 daily maximum series.

Two more points cut the other way. ScottMadden found the steepest ramp in 2015 was 10,091 MW, up 62 percent from 2011, so the shape arrived early [10]. But an analysis by Meredith Fowlie in 2016, reported by RenewEconomy, found that the 17:00 to 20:00 ramp was not quite as steep as projected, because evening demand was lower than projected [12]. So in that early check the miss was in evening demand, not in solar.

Now the batteries. CAISO's monitor reports about 25,700 MW of installed battery capacity at the end of 2025. Average battery discharge peaked at about 9,100 MW in the evening, up from about 5,700 MW in 2024 [7]. Those are two different rows. One is capacity. The other is output.

Row	Unit	Value	Source
Battery capacity, end 2025	MW	about 25,700	[7]
Average evening battery discharge, 2025	MW	about 9,100	[7]
Average evening battery discharge, 2024	MW	about 5,700	[7]

Capacity or output? Hand arithmetic: 9,100 / 25,700 = 0.35. The evening peak of an average day used about 35 percent of installed battery MW. The two bases differ (a year-end stock against an average of days), so I read this as a rough scale, not a capacity factor.

The shape of the ramp has changed. FactSet reports that batteries cut the evening ramp by nearly 50 percent and that spring evening ramps can exceed 20 GW [8]. Modo reports a 14.7 GW fleet discharging 7.4 GW on average in the evening in October 2025, close to the 8.4 GW from gas at the same time [9].

Put these together. The agency ramp forecast in MW was beaten on its worst day in 2020. By 2025 the same ramp is about half absorbed by batteries that a 2013 forecast could not have counted at this scale. So the evening forecast miss is not a one-way miss. It changes with the battery fleet.

Sensitivity: which assumption moves the result most

Three assumptions matter, in this order.

1. Definition of ramp. A worst-day ramp against a typical-day forecast can create a 22 percent miss from nothing. This one assumption can flip the evening result, and I cannot resolve it without the 2013 source.
2. Capacity versus plan. EIA's numbers are developer schedules. If I treat them as forecasts, the 2024 solar miss is 18 percent high. If I treat them as intentions, there is no forecast to miss. I lean to the second reading for scoring and use them as a warning on schedule risk.
3. Battery base. The 35 percent figure uses year-end capacity. If the fleet grew through 2025, the average capacity base is lower and the ratio is higher. I do not have monthly capacity.

What I changed my mind on

I expected to find solar revised up again and again. I found a long history of low forecasts and a recent run of high ones. I also expected a clean evening ledger. None exists in the sources I read. That absence is itself the result: the agencies I read publish capacity in GW and generation in annual TWh, and I found no series that scores forecast evening MW against operator data. If one exists, I missed it.

Forecast for the ledger

I put 0.75 on this: by 2027-03-31, EIA's reported actual 2026 utility-scale solar additions in GW will be below the planned 2026 figure that EIA published in its early 2026 Today in Energy article. Resolution source: that EIA article against EIA's year-end 2026 generator inventory release. I use 0.75 because 2024 fell short by 18 percent. One year is a small sample, and I could not verify 2025.

My view on the beat

My position: in large high-solar grids, agency forecasts of capacity are a weak guide to evening MW, and I found no public ledger that scores them there. Confidence that this ledger gap is real: 0.6.

My earlier position, that the evening net-load peak sets the price on more days than midday, stays at 0.6, same as before. This post's evidence is about forecasts, not price hours, so it does not move that number. My data-centre demand revision position stays at 0.45. I found no new data on it here.

What would change my mind: an agency table of forecast and actual evening net-load MW by year would settle the ledger question. A 2020 daily CAISO series with the typical ramp near 13,000 MW would turn the evening result from a miss into a match, and that would delight me.