1. How can LAMEPS * help you to make a better forecast for extreme weather Henrik Feddersen, DMI hf@dmi.dk * LAMEPS =Limited-Area Model Ensemble Prediction System

2. Outline Motivation Motivation Ensemble methodology Ensemble methodology Rainfall case studies Rainfall case studies Probability upscaling Probability upscaling Verification, summer 2011 Verification, summer 2011 Wind case study Wind case study Summary Summary

3. Motivation Assess forecast uncertainty Assess forecast uncertainty Assess risk of high-impact weather Assess risk of high-impact weather Heavy rain (>24mm/6h) Heavy rain (>24mm/6h) Cloudburst (>15mm/30min) Cloudburst (>15mm/30min) Heavy snowfall (>15mm/6h) Heavy snowfall (>15mm/6h) Snowstorm (>10mm/6h and >10m/s) Snowstorm (>10mm/6h and >10m/s) Storm (mean and gust) (>24m/s) Storm (mean and gust) (>24m/s) Hurricane (mean and gust) (>32m/s) Hurricane (mean and gust) (>32m/s)

4. Ensemble methodology Sample uncertainty in Forecast initial conditions Model formulation Lateral boundary conditions

5. Initial conditions Perturbation of analysis ObservationsAnalysisObservationsForecast Perturbed analyses Ensemble members

6. Initial conditions Perturbation of observations ObservationsAnalysisObservationsForecast Ensemble data assimilation Ensemble members Perturbed observations

7. Model uncertainty Multi-model ensemble NWP Model A Model B Model C

8. Model uncertainty Stochastic physics InitializationDynamics Postprocessing Physics NWP model

9. Model uncertainty Multi-scheme ensemble TurbulenceSurface 1 Surface 2 Radiation Convection 1 Convection 2 NWP model physics

10. Model uncertainty Multi-parameter ensemble TurbulenceSurface Radiation NWP model physics Convection

11. Limited-area ensembles vs global ensembles Global ensembles Uncertainty in synoptic development in the medium- range Limited-area ensembles Uncertainty in mesoscale development in the short- range

12. DMI-HIRLAM Ensemble Prediction System Resolution = 0.05° horizontal / 40 vertical levels Resolution = 0.05° horizontal / 40 vertical levels Members = 25 Members = 25 Forecast length = 54h Forecast length = 54h Forecast frequency = 4 times per day Forecast frequency = 4 times per day Initial and lateral boundary conditions = 5 Initial and lateral boundary conditions = 5 Scaled Lagged Average Forecast (SLAF) error perturbations Scaled Lagged Average Forecast (SLAF) error perturbations Cloud schemes = 2 Cloud schemes = 2 STRACO and KF/RK STRACO and KF/RK Stochastic physics = yes/no Stochastic physics = yes/no Surface schemes = 2 Surface schemes = 2 ISBA and ISBA/Newsnow ISBA and ISBA/Newsnow Independent of ECMWF's ensemble prediction system Independent of ECMWF's ensemble prediction system

13. Short-range ensemble spread

15. Case study, 2 July 2011 Precipitation stamp map

16. Case study, 2 July 2011 Probability map Probability = 10-20%: Only 3-4 members predict the event!?

17. Case study, 2 July 2011 50 mm/6h contours More than 4 members predict the event! Different members in different colours

18. Probability upscaling Conventional probability Conventional probability — In every grid point: Fraction of members that predict the event Upscaled probability Upscaled probability — In every grid point: Fraction of members that predict the event in a neighbourhood of the grid point — Probability that the event will happen somewhere near grid point

19. Probability upscaling example Prob = 1/25 Prob = 3/25 Prob = 8/25

20. Upscaled probabilities Max probability > 40% Upscaling diameter = 15 grid cells ~ 80 km

21. Verification of 2 July 2011 case Note the agreement between locations of max probability and max observed rainfall! Observed

22. Alternative verification

23. Location of ensemble member maxima Where is max precip most likely? Where is max precip most likely? Where density of ensemble members is highest! Where density of ensemble members is highest! Upscaling method will show just that! Upscaling method will show just that!

24. Heavy rainfall examples

25. Some things to consider... At what probability threshold should you take action? At what probability threshold should you take action? How does forecast skill depend on forecast range? How does forecast skill depend on forecast range? How many false alarms can you expect? How many false alarms can you expect?

26. Verification, JJA 2011 Relative operating characteristic Hit rate = events correctly foreacast / events occurred False alarm rate = events falsely foreacast / events non-occurred Perfect forecast (FAR,HR) if forecast, when prob > 1/25 (FAR,HR) if forecast, when prob > 2/25

27. Relative operating characteristic Upscaling vs No upscaling NB. False alarms are acceptable, if they are accompanied by nearby hits for the same forecast!

28. Relative operating characteristic Forecast skill as a function of lead time

29. Relative operating characteristic Ensemble vs Deterministic

30. Threat score If probability threshold = 50%... False alarm Hit Miss

31. Simplified threat score HIT = 1 if at least one hit HIT = 1 if at least one hit FA = 1 if at least one false alarm and no hits FA = 1 if at least one false alarm and no hits MISS = 1 if at least one miss MISS = 1 if at least one miss Count for each forecast... Purpose: Find optimal probability threshold for which an event should be forecast

32. Simplified threat score The threat score is maximized if warnings are issued when the forecast probability ≥ 20%

33. Simplified threat score The threat score is maximized if warnings are only issued when the max forecast probability > 45% Issue a warning only if the max forecast probability exceeds a certain threshold (will reduce false alarms)

34. Guidelines to forecasters 20-50% probability: Pay attention 20-50% probability: Pay attention > 50% probability: Take action > 50% probability: Take action

35. Why not use ECMWF's ensemble prediction system? LAM-EPS 50mm contours ECMWF-EPS 50mm contours

36. Why not use ECMWF's ensemble prediction system? LAM-EPS 50mm contours ECMWF-EPS 15mm contours 50mm/10km 2 vs 15mm/1000km 2

37. Wind case, 8 Feb 2011 Deterministic model vs ensemble mean 24h forecast

38. Wind case, 8 Feb 2011 Deterministic model vs ensemble mean 12h forecast

39. Wind case, 8 Feb 2011 Deterministic model “Truth”

40. Summary Limited-area, high-resolution, short-range ensemble forecasts can provide guidance for extreme weather for localized events Limited-area, high-resolution, short-range ensemble forecasts can provide guidance for extreme weather for localized events Particularly useful for forecasting the location of extreme events, e.g. convective rainfall events (using the upscaling method) Particularly useful for forecasting the location of extreme events, e.g. convective rainfall events (using the upscaling method) Guidelines must be provided for the usage of probabilistic forecasts of extreme events, e.g. Guidelines must be provided for the usage of probabilistic forecasts of extreme events, e.g. 20-50% probability: pay attention 20-50% probability: pay attention > 50% probability: take action > 50% probability: take action Upscaled probability forecasts have been used frequently by DMI forecasters for guidance this summer Upscaled probability forecasts have been used frequently by DMI forecasters for guidance this summer Focus on rainfall events, but also potential for wind events Focus on rainfall events, but also potential for wind events