Repository navigation
PV calibration has been reworked. It now learns from 28 days of stored evidence rather than the last 7 days of history, and it only corrects your forecast as far as that evidence supports. A run of dull days no longer drags down the next sunny day's forecast, and forecasts that a provider has already calibrated for your site are mostly left alone. A steady bias, such as shading, a misconfigured array or inverter losses, is still corrected.
What's Changed
New features
- feat(solar): PV calibration from a 28-day evidence store with self-limiting factors by @springfall2008 in #6231
- 28 days of evidence, kept in storage. Each day's forecast and actual generation is saved per half-hour slot. The first scan reads up to 28 days of history. After that it reads 2 days once a day, or more after a gap.
- Factors that move only as far as the evidence supports. The day factor and the half-hour shape are each pulled back toward 1 by their own day-to-day noise. Recent days count for more: a day's weight halves every 3.5 days.
- New PVCalibration chart page, replacing PVAccuracy. It shows a status table, each stored day's actual ÷ forecast ratio by slot against the factors applied, and a summary in two-hour bands. Old
chart=PVAccuracylinks still open it. - Backtest results: on a raw Forecast.Solar site, the day error on sunny days after 3 dull days fell from 18.4% to 7.9%. Day error on sites whose forecast was already calibrated went from 20.6% with the old method to 19.0%, against 19.3% with no calibration.
- New optional
apps.yamlsettings:pv_calibration_history_days(28),pv_calibration_scan_days(2),pv_calibration_half_life(3.5),pv_calibration_shrink(0.20) andpv_calibration_shrink_slot(0.20). They are documented in apps.yaml settings. - Unchanged: the P10/P90 band and the array cap's observed peak still use the last 7 days.
days_use_scaling,calibrate_band, clip imputation andmetric_pv_calibration_enablebehave as before. - Fixed along the way: calibration placed generation 5 minutes early. The oldest day in the window could also be counted with only half of its forecast.
Bug fixes
- fix(solar): PV calibration follow-ups - restarts, clock changes, cheaper retries, page labels by @springfall2008 in #6233
- A restart now reads only the days since the last scan, not all 28 days again.
- If the forecast sensor has no recorded history, the scan stops before the long generation history read, so hourly retries are cheap.
- Half-hour slots stay in local time across clock changes, and a rescan runs when the clock moves back.
- The array cap's observed peak skips down days again.
- The PVCalibration page labels have been tidied: the next scan time, the day factor's spread as a percentage, and default day counts before the first run.
- fix(web): Curves page captions wrap by @springfall2008 in #6218
Full Changelog: v9.8.1...v9.9.0