Probabilistic forecasting of (severe) thunderstorms in the Netherlands using model output statistics

1
Probabilistic forecasting of (severe)
thunderstorms in the Netherlands using model
output statistics
Photo R. Hoetink (17 July 2004)

• Maurice Schmeits, Kees Kok and Daan Vogelezang
• Royal Netherlands Meteorological Institute (KNMI)
• Department RD Models

2
Outline

• Model output statistics (MOS)
• Data and predictand definitions
• Illustration of logistic regression
• Case 17 July 2004
• Verification results
• Conclusions plans

3
Model output statistics (MOS)

• Goal
• Features

To determine a statistical relationship (mostly
via regression) between a predictand (i.e. the
occurrence of a thunderstorm in this case) and
predictors from NWP model forecasts

• separate regression equation for each projection
• correction of systematic model errors as a
  function of projection

4
Predictand definitions (from the SAFIR lightning
dataset)
Predictand for thunderstorms Probability of gt
2 lightning discharges in a 6h period (03-09,
09-15, 15-21 or 21-03 UTC) in a 90x80 km
region. Predictand for severe thunderstorms
Conditional probability of gt 500 discharges in a
6h period in a 90x80 km region, under the
condition that gt 2 discharges will be detected
in the same period in the same region.
5
MOS system for (severe) thunderstorms

• potential predictor set 1 15 thunderstorm
  indices, computed from 55 or 22-km HiRLAM
  forecasts
• potential predictor set 2 ECMWF (derived) DMO
  (forecasts)
• potential predictor set 3 (co)sinusday of the
  year
• developmental dataset warm half year
  (16/4-15/10) 1999-2002
• verification dataset warm season (19/5-7/10)
  2003
• severe thunderstorms 6 land and 6 coastal
  regions pooled separately (09-21 UTC) or all
  regions pooled (21-09 UTC)
• regression equations contain at least 1 and at
  most 5 predictors
• the system runs after each HiRLAM run (i.e. 4
  times a day)

6
Example of logistic regression equation using
only the first predictor (region E-MN period
09-15 UTC)
Probability of thunderstorms
crosses binary predictand triangles conditional
mean pluses logistic curve
Square root of convective precipitation sum from
12-18 UTC ((0.1 mm)1/2) EC12 30h
7
Conditional probability of severe thunderstorms
as a function of Boyden or CAPE (land regions
period 09-15 UTC)
crosses binary predictand triangles conditional
mean pluses logistic curve
Cond. prob. of severe thunderstorms
Cond. prob. of severe thunderstorms
Boyden index H00 12h
CAPE (J/kg) H00 12h
Boyden 0.1(z700 z1000) T700 200
8
Case 17 July 2004 (15-21 UTC)
SAFIR no. of discharges (n) (1089 lt regional n
lt 15532)
06 UTC run (based on H 1706 and EC 1612)
(source I. Holleman, KNMI)
Probability of thunderstorms
Cond. prob. of severe thunderstorms
Clim. prob. of thunderstorms 8-25 Clim. prob.
of severe thunderstorms 1-5
9
Verification results 2003 (Probability of gt 2
discharges 00 UTC run) Brier skill score
Brier skill score ()
Projection (hours)
10
Verification results 2003 (Cond. prob. of gt 500
discharges 00 UTC run) Brier skill score
Brier skill score ()
triangles all regions crosses land regions
diamonds coastal regions
Projection (hours)
11
Conclusions plans

• MOS equations for thunderstorms (gt 2 discharges)
  are skilful, possibly with the exception of the
  night (21-03 UTC) and especially the morning
  (03-09 UTC) in the coastal regions.
• MOS equations for severe thunderstorms (gt 500
  discharges) are skilful, with the exception of
  the morning in all regions and the afternoon
  (09-15 UTC) in the coastal regions.
• (Slightly) positive bias for nearly all
  projections and nearly all regions, so in general
  there is (some) overforecasting .
• Plan RD of a skilful(!?) method for short-range
  (0-12h) probabilistic forecasting of even more
  extreme convective events (using MOS and logistic
  regression/(non)parametric density estimation).
• Wea. Forecasting preprint www.knmi.nl/schmeits/T
  hunder_prep.pdf