The model
Three global ensemble models, one local truth. Every forecast cycle the system extracts ECMWF IFS ENS, ECMWF AIFS ENS and DWD ICON-EU at the orchard, then:
- calibrates each model's mean and spread against the on-site station, per lead time — fitted by maximum likelihood and validated out-of-sample (fit one season, verify the next);
- blends them by verified likelihood on identical samples, tempered to the effective number of nights so no model is silenced on noise;
- decides: one probability distribution per half-hour of the night → frost probability, expected hours below the critical temperature, alert threshold set by your protection economics.
Loading...
Simpler won on evidence: a 4-parameter calibration beats a 54-parameter one out-of-sample at every lead, the 51-member ensembles beat the high-resolution deterministic model even at short range, and calibration never hurts (hard-checked in the pipeline).
Static snapshot of the Phase-1 back-analysis; full methods and math live in the repository's back-analysis report.