Priority project Advanced interpretation COSMO General Meeting, 18. September 2006 Pierre Eckert
Recognition of high impact weather boosting method for thunderstorm prediction Initialisation of forecast matrix use either MOS on global models or DMO from LM Gridpoint statistics neighbourhood method Hydrological Applications applications with COSMO LEPS (MAP D-PHASE)
Automatic Weather Interpretation using Boosting Monday, September 18, 2006 COSMO General Meeting 2006, WG4 Donat Perler (ETH Zürich) Oliver Marchand (MeteoSwiss)
4 Supervised Learning Historic Data (a) Input Data (Model Output) (b) Label Data (SYNOP & lightning data) Learner Classifier New Data yes/no
5 Average final scores for 5-fold cross validation for the whole year 2005 ClassifierPODFARFBICSIHSS DWD (optimized for DE) 18%94% DWD (optimized for CH) 45%68% AdaBoost.M1 (DWD features) 57%59% AdaBoost.M1 (51 features) 72%34% Linear Discriminant (51 features) 57%58%
6 Operational Implementation of Boosting Example: 11 August 2006
7 Lightning data indicate thunderstorm in northeastern Switzerland
8 3h aLMo sums of precipitation for the same period show no signal!
Lokal-Modell Kürzestfrist Kürzestfrist = very short range (< 18 h) gridbox size: 2,8 km developed at DWD (Baldauf, Seifert, Förstner, Reinhardt, Lenz, Prohl, Stephan, Klink, Schraff) pre-operational since late summer 2006 LME GME LMK What is LMK?
What is „Neighbourhood Method“? Aims: account for general predictability limits in LMK output interpret small scales of LMK output statistically derive probabilistic forecasts from a single simulation Method: statistical post-processing spatio-temporal neighbourhood around each grid point derive pseudo-ensemble Application: surface fields of LMK output (Hoffmann, COSMO Newsletter No.6)
13 elements have been covered so far: 2m-temperature below freezing point wind gusts exceeding certain thresholds (14 m/s, 18 m/s, 25 m/s, 29 m/s, 39 m/s) rain amount exceeding certain thresholds (10 mm/h, 25 mm/h) thunderstorm (3 categories of severity) black ice New Focus: Warning Events
% probability of thunderstorm occurence from the neighbourhood method Example for Thunderstorm Prediction 25 June UTC + 18 h LMK test suite 3.3d
13 Shape of the neighborhood (P. Kaufmann) cylindrical rather than ellipsoidal independent spatial and temporal uncertainty true for no or weak advection, wrong for strong advection x y t
14 Linearly fading weights Circles around singular high model values too well visible Idea: smoother edges Introduce linear fading of weights (relaxation) Adds sponge layer around cylindrical neighborhood large, small neighborhood
:00 UTC moderate prob. – event occurred 50 mm / 24 h
:00 UTC raw model output 50 mm / 24 h
17 Neighborhood method Combination of Ensemble and Neighborhood method would combine both synoptic-scale and small-scale uncertainties
18 Plans for next year The weight of the project will be displaced on the verification of very high resolution models, mainly precipitation Proposed verification methods always use some aggregation on gridpoints The optimisation of the aggregation is using the verification WG4-WG5 project
19 Radar12 km forecast1 km forecast mm The problem we face 0100 km Six hour accumulations 10 to 16 UTC 13th May 2003 From N. Roberts, UKMO