Most solar forecasts answer the wrong question. They give you a single number for 2pm tomorrow — and that number is wrong by definition, because cloud cover is a probability, not a fact. An operator scheduling load, a trader managing an imbalance position, a plant avoiding a deviation penalty: none of them can act on a point estimate. They need to know the range and how much to trust it.

So Vastvic Solar forecasts distributions, not points. For each of the next 96 quarter-hours — a full day ahead at 15-minute resolution — it predicts a low, a middle and a high: P10, P50 and P90. The band is the product. The middle is just the part that fits in a headline.

A forecast without a range tells you what to hope for. A forecast with one tells you what to plan for.

Physics the model can feel

Generation isn’t a mystery time-series; it’s driven by the sun. The feature set encodes that: cyclical time-of-day and season, clear-sky irradiance and solar elevation from first principles, a clear-sky index that isolates the cloud signal from the daily arc, lags at multiple scales, rolling statistics, and weather inputs. The model learns how far reality strays from clear-sky physics — which is exactly where the uncertainty lives.

Quantiles, kept honest

Three LightGBM quantile models predict P10, P50 and P90, with post-processing that guarantees they stay ordered and non-negative — no forecast where the “low” sits above the “high.” For plants with the data to justify it, an optional attention-based deep model (trained with pinball loss) can take over the same contract.

Earning the right to ship

Every model is scored on metrics that mean something for solar: skill score against a persistence baseline (does it beat “tomorrow looks like today”?), MAE and MAPE measured only during generation hours, and interval coverage — do the P10–P90 bands actually contain reality the right fraction of the time? A release gate then enforces two things before promotion: the model must beat persistence, and its intervals must be calibrated (coverage in a sensible band, neither over- nor under-confident). Miss either, and it doesn’t ship.

The hard part of forecasting isn’t producing a number; it’s producing a number you’re willing to stand behind. Quantiles give operations a range to plan against, skill score proves the model is worth more than a naive guess, and the release gate makes “trust the forecast” a property you can check — not a slogan.

Key takeaways

  • Forecasts are calibrated P10/P50/P90 bands, 96 steps of day-ahead at 15-minute resolution.
  • Solar-physics features (clear-sky irradiance, elevation, clear-sky index) isolate the cloud signal.
  • Models are judged by skill score vs persistence and by interval coverage — not point error alone.
  • A release gate blocks any model that can’t beat persistence or whose intervals aren’t calibrated.