Simulate or remember? What the forecasting literature says about PV simulation vs. site history

[HuggeK]

The claim on the table, raised while discussing how far to trust the planner's
PV input: simulating the PV panels — roof geometry, orientation, irradiance —
can be on par with just using the site's historic production data. With #1030
having just made Core the planning champion, whether the simulation side of
the forecast can carry real weight is worth grounding. This is what the
forecasting literature actually says (web research 2026-09-02, sources at the
bottom; every figure carries its paper).

The verdict, by situation

• Holds at day-ahead — the horizon an MPC planning against spot prices
  cares about. Day-ahead error is dominated by the weather forecast, not the
  PV conversion: a well-parameterized physical chain and an ML model fed the
  same NWP land within ~5–15 % of each other, and the strongest head-to-head
  found (Mayer 2022) had the optimized physical chain slightly ahead of pure
  ML trained on two years of history. Visser 2022 found the physical model
  delivered the best economic value on the day-ahead market.
• Holds for cold start. For a site with no history, physical simulation is
  the accepted answer — stated explicitly in the literature and shipped by
  industry (Solcast forecasts from location + capacity alone, then recommends
  parameter tuning once ~3 weeks of measured data exist).
• Fails intra-hour and intra-day. Below roughly 4–6 hours, persistence and
  statistical models on recent measured output beat NWP-driven simulation.
  That is history's home turf.
• Fails on whatever the static picture misses. Degradation (~0.5 %/yr),
  soiling, inverter clipping, curtailment — and in the Nordics, snow: 1–20 %
  annual loss, up to 100 % in a winter month. All invisible to geometry, all
  absorbed automatically by history. Snow alone can exceed the entire
  shading-model error budget in a Swedish winter.
• Backbone for net-meter decomposition. With no dedicated PV meter there
  is no clean history to learn from — the best behind-the-meter
  disaggregation methods all embed a physical PV model (44 % MSE reduction
  over purely statistical state of the art in the flagship IEEE study).

What a LiDAR-grade simulation actually buys

Validated LiDAR-plus-simulation studies land within ±4–8 % on annual
energy (best case 3.6–5.3 % with measured weather — Jakubiec & Reinhart,
MIT/Cambridge). What it recovers is the static geometric part of site
specificity — real azimuth, tilt and horizon shading — which is precisely the
part generic satellite-tilt tools miss. What it cannot recover is anything
that changes after the scan: the point cloud is a snapshot, and vegetation
grows.

How fast does history catch up?

Slower than intuition says, at day-ahead. The cited minimum for ML to learn
seasonality is one calendar year; in the strongest comparison found, even
two years of history did not let pure ML beat an optimized physical chain.
But history pays off much earlier — from about three weeks — not by
replacing the simulation, by tuning it: bias correction, parameter
refinement. Hybrids beat both parents in essentially every comparison in a
180-paper systematic review.

Does forecast accuracy even move the battery?

Less than the argument assumes. MPC studies show battery capacity cushions
forecast error: controllers with imperfect forecasts still reach 96–98 % of
optimal dispatch value, and weather-forecast error costs only 0.5–4.5
percentage points of self-consumption. What hurts an MPC is not noise but
persistent bias — wrong azimuth, unmodeled shading, snow — which is
exactly what simulation (geometry) and history (correction) each fix half of.

The consensus architecture — which FTW already has the bones of

Physical model as prior, statistical correction on top, weights shifting
toward data as it accumulates. Unambiguous across academia (IEA PVPS
Task 16, Mayer, the systematic reviews) and industry (Solcast).

Mapped onto the codebase: the roof model supplies geometry on day one
(#735) · the plane-of-array projection turns forecast irradiance into
per-array watts (#718, merged) · STRÅNG scores expected-vs-actual and feeds a
guarded calibration factor back (#734) · the RLS twin keeps learning the
residual (go/internal/pvmodel). That is the hybrid — the pieces just need
to keep talking to each other. Nothing in the literature asks the planner to
sit closer to the PV model than a forecast contract: the value arrives
through the input, and the day-ahead accuracy gap between forecast families
is small next to the bias problems the hybrid solves.

Numbers to keep in the room

Fact	Figure	Source
STRÅNG GHI vs 12 Swedish stations	~30 % hourly · 16 % daily · 8.9 % monthly RMSE	SMHI validation
LiDAR sim vs metered annual energy	3.6–5.3 % (best) · ±7.7 % over 886 homes	Jakubiec & Reinhart; OSTI
Nordic snow losses	1–20 % annual · ≤100 % monthly	RSER 2025; FME Solar
Module degradation	~0.5 %/yr median	Jordan & Kurtz, NREL
Hybrid vs its parents (day-ahead MAE)	−5.2 % vs physical · −10.4 % vs pure ML	Mayer 2022
MPC value with imperfect forecasts	96–98 % of optimal	Berkeley ACC15
BTM disaggregation with a physical PV model	−44 % MSE vs statistical SOTA	Kabir et al., IEEE

Metric trap: solar-forecast errors are normalized three different ways —
installed capacity, mean production, daytime-only — which differ by a factor
of 3–5. The widely-quoted "ML gets 5–9 % vs physical 40–50 %" mixes
denominators and should not be used as evidence either way.

Sources

• Mayer & Gróf (2021), Extensive comparison of physical models for PV power forecasting, Applied Energy
• Mayer (2022), Benefits of physical and machine learning hybridization, RSER
• Visser, AlSkaif & van Sark (2022), Operational day-ahead solar power forecasting, Renewable Energy
• Nguyen & Müsgens (2021), What drives the accuracy of PV output forecasts? — 180-paper systematic review
• IEA PVPS Task 16 (2020), Regional PV power forecasting benchmark
• Jakubiec & Reinhart (2013), LiDAR + simulation vs measured rooftop production, Solar Energy
• Lingfors, Killinger, Engerer, Widén & Bright (2018), LiDAR-derived vs output-inferred shading, Solar Energy
• RSER (2025), Snow losses on Nordic PV — review · FME Solar modelling recommendations
• Jordan & Kurtz (NREL), PV degradation rates — analytical review
• Kabir, Duan et al., Behind-the-meter solar estimation, physical + statistical, IEEE
• AutoPV (2022), Cold-start forecasting from fleet-pretrained ensembles
• SMHI STRÅNG validation · Solcast Advanced PV power model · Berkeley ACC15 MPC study

Figures read from search extracts rather than full paywalled texts were
sanity-checked for consistency but not re-derived; treat exact decimals from
the paywalled rows with mild caution.