New Precipitation Scheme in the mLM
meteoblue has introduced a new precipitation scheme in the meteoblue Learning MultiModel (mLM). It delivers up to 9 % higher forecast skill than leading global weather models, holds that gain where there are no stations, and switches less between rain and no rain.
What changed
Precipitation from the mLM in the Forecast API is now calculated with a new scheme. The mLM combines the output of several weather models. Under the previous scheme it usually ranked among the top three, and a different model came out ahead each month depending on the region. The new scheme scores higher than leading global weather models, by up to 9 % in Heidke Skill Score (HSS).
Requests, variable names, units and packages stay the same, so no change is needed on the customer side. The same forecast is used on the meteoblue website and in the meteoblue apps.
What improves
- Skill. Up to 9 % higher HSS across precipitation thresholds. HSS measures how often a forecast gets rain and no rain right, corrected for the hits expected by chance.
- Regions without stations. The improvement is similar to that in well-observed regions.
- Stability. Forecasts switch less between rain and no rain when new model runs arrive.
- Consistency with probability. When precipitation probability is high, the forecast shows rain more often.
The scheme is tuned for a low false-alarm rate together with a good probability of detection. A model that forecasts rain often will catch almost every event, but the false alarms that come with it make the forecast harder to plan with.
Where it applies
The new scheme applies globally, except in the tropical band between 25°N and 30°S. There, the mLM keeps the precipitation scheme in use since 2018, and forecasts do not change.
Known limitations
Heavy precipitation is not yet fully tuned, and further improvement is expected for the highest thresholds.
Verification
Verification reports for the new scheme will be published alongside meteoblue's existing forecast verification reports in the future.