Model accuracy: Denver example

• Sun hours: Golden (NLR Solar Radiation Research Laboratory) logs sunlight every minute. Nov 5: 1,440 readings add up to 358,985 W/m²·min; ÷ 60 ÷ 1,000 = 5.98 kWh/m², which is 5.98 sun hours. Sensor: 90-South PSP.
• Wh per sun hour: the Designer (Fit 1, v323, typical-year weather) expects 21.3 kWh in November = 711 Wh a day, from 5.20 sun hours a day on the panel with open sky: 711 ÷ 5.20 = 136.9. October: 598 ÷ 4.78 = 125.1.
• Old baseline, same steps with the Oct 28 report (21 kWh in November, 5.20 sun hours): 134.6 Wh per sun hour × 5.98 = 806 Wh.
• Limits: temperature and wind are typical-year, not that day’s; Golden has open sky and no building shade.

The Designer’s estimate rests on a few things set by hand: which way the panel faces, its tilt, and how the site photo is lined up with the compass. A small slip in any of them moves the shade edges by many minutes and the day’s energy by several percent. The fit checks those settings against what the panel actually did.

1. Answer key. A cloudless day from the inverter’s own export (Nov 5 full day, Oct 29 morning), plus shade-edge times read off the app charts (Oct 28, Nov 7). Clouds would hide the geometry, so only clear days are used.
2. Every combination. Five knobs, each swept across its range: facing (±10°), tilt (80–90°), photo rotation (±5°), photo level (±2°) and photo width (±5 %). A coarse grid first, then a fine grid around each leading answer: 75,330 runs of the Designer’s own clear-day model, about 10 minutes.
3. Score. Each run’s curve is compared with the measured curve minute by minute (RMS error, in watts), with the misses at the shade edges (in minutes) counted in. Lower is better.
4. Top three. Near-identical combinations are grouped, and the three best distinct answers are shown below. An answer is listed only if it scores within 15 % of the best.

What it determines

• Was the panel installed as designed? Facing and tilt. Here all three answers say vertical, facing 176–180°, so yes.
• Is the photo lined up right? Rotation, level and width set when the shade arrives and leaves. Fit 1 moved the morning edge from about 20 minutes off to within 2.
• What is left over? There is no catch-all “fudge” knob. Whatever the five knobs can’t explain (about 2 % here) is reported as a gap, and it tests the loss and inverter assumptions.

How it is used

• The best answer, Fit 1, became the site’s settings. It drives the updated Denver report and the daily estimates in the accuracy log (sun hours × Wh per sun hour).
• The original design stays on record as the old baseline, so the gain is visible.
• For the tool, it shows which inputs matter most. Photo alignment matters more than panel facing, which is where the Designer’s photo step should guide users.

• The panel is where the design says: vertical, facing 176–180°. All three answers agree; the facing can’t be pinned closer than that yet.
• The old 80.5° lean-back is gone. The Oct 29 fit leaned the panel back to make up for a clear-sky model that ran low at Denver’s altitude. With the altitude correction in the tool, the vertical panel fits as built.
• Tilt sits on the sweep limit (90°). The range stops at vertical; a phone level on the panel face would close this out.
• Remaining gap about 2 % high: inside the panel’s ±5 % nameplate tolerance; no knob absorbs it. Air temperature is still a flat 20 °C (see Method).
• Photo knobs settled together: rotation +2°, level +1°, width ×0.96, now that three edges pin them.

75,330 combinations, coarse sweep then a fine sweep around each leader, about 10 minutes.

• Morning edge (vertical building edge) checks the photo’s rotation; evening edge (near-flat roofline) checks its level.
• All by-eye edges land within their reading error (±4 to ±5 min).

Golden’s 1-minute readings averaged to the inverter’s 5-minute steps, scaled by the 180 W nameplate (1,000 W/m² = 180 W), on the same clock (MST). The gap between gold and green is everything between sunlight and the outlet: heat, wiring, inverter, glass angle, and the balcony’s shade.

• Clear days agree: at midday the balcony delivers 0.77–0.79 of Golden’s sunlight × nameplate, three different days. Oct 28’s 0.71 is its clouds.
• Source curves: Oct 29 and Nov 5 from the maker’s web chart; Oct 28 and Nov 7 read off app screenshots. The 5-minute CSV export would replace all four with exact values.

Cloud lag, Golden to balcony

• Not measurable yet. Each day’s Golden curve was slid ±60 min against the balcony’s (slow trends removed, so only cloud dips count). No shift matches well (r 0.16–0.47) and the best shifts disagree (−41 to +59 min): there’s nothing shared to line up.
• Clocks agree: on clear days the shade edges land within 1–2 min of the sun’s position, so no time correction is applied to the overlays.
• What would measure it: partly cloudy days where the same cloud field crosses both (Golden shows candidates on Oct 30–Nov 4) with the inverter’s 5-minute export. Each matched day adds one lag reading; the lag is applied once several agree.

133 Wh per Golden sun hour; Golden’s sunlight explains 95 % of the day-to-day swing (14 days; darker dots are clear days).

Each knob has a value, a range and a lock. Locked knobs stay put; the sweep moves the rest. Only things that can be checked on site.

• Result: up to three distinct answers; near-identical combinations are grouped; one answer when the rest score more than 15 % worse.
• No catch-all efficiency or scale knob. What the knobs can’t explain is reported as the energy gap.

Settings, not swept. Defaults are the Designer’s PVWatts values; change one only with a checkable reason.

Default: PVWatts part-load curve at 96 % nominal. Load is a share of the inverter’s DC rating (360 VA ÷ 0.96 = 375 W).

• This site runs at part load: a 180 W panel peaks near 32 % of the inverter’s rating, so most of the day sits on the steep part of the curve.
• Next: replace the curve with measured efficiency by power range (e.g. CEC test points for this inverter) when available.

• Model: the Designer’s own clear-day model and sun math, run unchanged for every combination.
• Score: RMS of measured minus model, 5-minute points, plus shade-edge timing from other days.
• Sweep: coarse grid over unlocked knobs, then a fine grid around each leading answer.
• Altitude: clear-sky beam corrected for site elevation (Laue 1970), the same correction now in the Designer.
• Air temperature: 20 °C assumed until hourly weather is loaded.