github springfall2008/batpred v9.9.0
PV calibration learns from 28 days of evidence and only corrects a steady bias

43 minutes ago

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=PVAccuracy links 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.yaml settings: pv_calibration_history_days (28), pv_calibration_scan_days (2), pv_calibration_half_life (3.5), pv_calibration_shrink (0.20) and pv_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 and metric_pv_calibration_enable behave 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

Don't miss a new batpred release

NewReleases is sending notifications on new releases.