Adopted and open questions

No experiment result is adopted in production, a set of results are candidates for the next model release, and each open question belongs to one theme. This page gives one line per item and links the theme that owns it. A planner checks here before proposing an experiment that a theme already holds as settled or planned. Evidence tags follow the update rule.

Adopted

• Nothing is adopted. The release log lists no numbered model release, and no result.md carries production_status: adopted (model releases) (docs).

Candidates for the next release

• Primaries and supply points: weekly refits, and a five-member blend at Primaries (experiment 87). Owner: production reference and best guess (docs).
• Retraining: weekly refitting without the 14-day correction (experiments 34, 48, 81). It holds at 16 Primaries on two sets of runs, but 3 series carry much of the gain. Owner: retraining and training history (exp 34, 48, 81; 16 Primaries; provisional).
• Solar and wind farms: the pooled mixture over ensemble members, the solar spread factor, and the wind power-curve base margin (experiments 2, 16, 20). Owner: calibration and uncertainty (docs).
• Solar offset: a forecast of the unmetered solar output as an XGBoost base margin at Primaries with a clean capacity estimate (experiments 64, 78). Owner: solar disaggregation (docs).
• Physics-based demand model: the convex model with a Kalman filter (experiment 111), a recent-error correction recipe (experiment 155), and a correction neural network with a two-number solar summary (experiment 220). Owner: differentiable physics: forecast engine (docs).

Open questions by theme

• Do the spread factor, the blend, and the solar offset hold on a second validation year? The first year chose their settings. Owner: production reference and best guess (docs).
• Which training source suits solar and wind farms, the ERA5-trained mixture or the ensemble's own history? Needs more than 6 solar and 3 wind farms. Owner: production reference and best guess (docs).
• Does weekly refitting beat the stack on an unscored year and with the full ensemble, and why do older training years score worse? Owner: retraining and training history (docs).
• Do old years of history help the physics-based demand model? The answer changed sign between two scored years. Owner: differentiable physics: long history and drift (docs).
• Does one model pooled across Primaries match a model per substation, and does pooling hold at hundreds of series? Owners: hierarchy and pooling and differentiable physics: pooling across substations (docs).
• Does a calibrator that sees recent power help solar and wind farms, and does a missing-lag flag widen the band where a lag is missing? Owner: calibration and uncertainty (docs).
• Does lagged power help once the 14-day correction is in place? No experiment has combined them. Owner: lagged power and error correction (docs).
• Does the live Met Office UKV feed give wind farms the gain the research archive did? Owners: generator forecasting and weather products and blending (docs).
• Does a better sunshine forecast close the gap to the true-sunshine oracle? Owner: solar disaggregation (docs).
• Which extra inputs help Primary demand beyond the calendar and solar set? A heating term gave 0.14 points in cold hours on one winter, under its 0.3 target. Owner: demand inputs (docs).
• Does any forecast curtail less than another once about 2,500 series are scored, and are the forecast tails too narrow in general? Owner: curtailment cost (docs).
• What physical plant reproduces each metered solar farm's output, and how much rooftop solar sits behind each substation? Owners: metered solar farms and differentiable physics: embedded generation (docs).
• Does the physics-based demand model, driven by ensemble weather, forecast Primary demand up to 7 days ahead as well as BG2? Experiment 111 could not tell them apart. Owner: differentiable physics: forecast engine (exp 111; 5 Primaries, one year; provisional).
• How should the physics-based demand model combine its daily profile with holiday, heating, and solar terms? Owner: differentiable physics: demand, calendar and weather (docs).

Last checked against experiment 225.