Sin ttulo de diapositiva

1
Verification of HIRLAM Surface SW Irradiance
Predictions using Pyranometer Data and Satellite
Retrievals
Carlos Geijo, Bartolomé Orfila INM (Instituto
Nacional de Meteorología), Leonardo Prieto
Castro, 7, 28037, Madrid, Spain.
BRIEF DESCRIPTION OF THE WORK DONE Recently we
have done some verification work at INM focusing
on the Surface Short-Wave Irradiance (SSWI)
parameter. The verification exercise comprised
three time periods with different climatological
characteristics (August 05, October 05 and
January 06). The geographical area of interest
for this work is the Iberian Peninsula, although
the pyranometer data comes only from Spanish
stations. The basic quantity that has been
compared is daily sums (or daily mean values if
expressed in W/m2). The HIRLAM model was run in
two different configurations or modes. A)
Hindcast mode, with HIRLAM version 6.4 using
European Centre analyses for initialisation and
boundary conditions to perform long range
integrations (up to D7). This setting was chosen
because these long integrations allow one to
explore predictability issues that are not
accessible with shorter range forecasts. B)
Forecast mode , with HIRLAM version 6.1.2 using
near-real-time 3D-Var HIRLAM analyses produced at
INM and European Centre forecasts for boundary
conditions. Only results for the winter period
are available in this second case. Both modes
were run at similar horizontal resolutions (0.2
degrees and 0.16 degrees respectively) and equal
vertical resolution.
VERIFICATION DATA USED Pyranometer data from
about twelve INM staffed stations evenly
distributed over Spain mainland, coast and
islands (Mallorca) were used for the comparisons.
These pyranometers have technical characteristics
and are operated in compliance with the WMO
secondary standard, the one with the highest
rank in its list of operational pyranometers. It
is however difficult to estimate the accuracy of
these measurements because of the numerous
effects affecting the functioning of these
thermoelectrical sensors. 5 relative error for
hourly sums is often found in the references as a
fair estimate for properly operated and
maintained equipment. Quality screening consisted
of rejection of those sites with gaps and/or
obvious errors (e..g. measurements out of
physical bounds) over the verification period.
After this data quality control was done, most of
the data were retained. Other source of data
used in this work was M8 (Meteosat 8) satellite
retrievals of SSWI fields. We have used daily
products from the CM-SAF (Satellite Application
Facility for Climate Monitoring). These fields
have nominal spatial resolution of 15 Km,
comparable to the spatial resolution of the model
output fields. Quality specifications for these
fields set the acceptance threshold at 10 W/m2
r.m.s.e on the monthly time scale. In the graphs
o the left of these lines we can see that, at
least for daily means, the pyranometer data and
the satellite data compare within acceptable
limits. For August 2005 they compare (not shown)
worse, but at that time the CM-SAF products were
still graded as pre-operational. That both
sources of information, in-situ and
remote-sensed, match so well raises the
confidence in the quality of the verification
data used in this work.
DESCRIPTION OF VERIFICATION RESULTS Hindcast
mode A) Comparison with Satellite Monthly Mean
Fields The maps beside these lines display
forecasted values averaged over the first 48
hours of integration (left), satellite
observations (centre) and the corresponding
difference fields (right) for August 05 (upper
row), October 05 (middle row) and January 06
(lower row). In the first case, the summer month,
the sharp gradient in the north of the Peninsula
is well reproduced (different spatial resolutions
of satellite fields and model fields can explain
the mismatch observed in the difference field for
such small structure). We can see too that the
model has been able to capture faithfully the
complex spatial patterns of the fields for
October and January, although in this latter case
frequent fogs in the Duero and Ebro basins
(arrows), which strongly modulate the SSWI field,
were not correctly simulated. The difference maps
indicate a conspicuous positive bias over land,
while over sea surface this bias is smaller or
even reverses the sign. This spatial signature in
the systematic error can be traced back to a
number of reasons residual cloudiness spin-up
effects, too high atmospheric transmissivity
values for clear-sky conditions (see below),
failure to simulate foggy weather, ...
B) Comparison with Pyranometer Daily Values The
time series (above this paragraph) give a first
idea about how well compare calculated and
in-situ measured values. They correspond to two
very distant sites and two different seasons of
the year. It is clear that the lines
(calculations) closely follow the marks
(observations),i.e, the system has, no doubt,
prediction skill. A closer look is provided by
the graphs (on the right of this paragraph ) that
show the error for D2 SSWI (i.e., mean
irradiance between 24 and 48 hours of
integration) versus measured value. The graphs
correspond to August (left), October (centre) and
January (right). We can see (August and October)
that the points lay above the x-axis on the
clear-sky end and present much more dispersion
on the cloudy-sky end. The thick lines indicate
the 100 relative error and the thin line the 50
relative error. For these two months the
clear-sky situations are responsible for the
overall positive bias (25 W/m2 and 15 W/m2,
respectively). In January (overall bias 8 W/m2)
the graph has a different aspect. One can notice
the bigger number of points off the 100 relative
error mark for situations with low measured
values. The following table summarises the
verification results (for August only results up
to D2 were available)
As far as systematic errors is concerned, it has
been found that the model calculations
overestimate pyranometer measurements by about
10 of the measured monthly mean value, for the
three periods of time. The bias does not show
dependency on the forecast range (no drift is
detected). Most of the bias appears under
clear-skies. Regarding random errors, the results
indicate an important modulation of their size
with climatological conditions, model and
pyranometer measurements compare worse for
January than for October or August. In these
later cases, the linear correlation coefficient
up to D3 is in the interval 0.8 to 0.7 while in
the former case its value drops more sharply
(0.73 to 0.56) with the forecast range. Typical
relative errors, with respect to measured monthly
mean, are in the 15 to 25 range for August and
October and 40 for January.
Forecast mode At the time of delivering
forecasts in a timely manner,some modifications
in the forecasting system configuration are
necessary with respect to the hindcast mode. In
order to assess how these modifications (see
Brief Description of the Work Done) impact on
the forecast quality, the verification exercise
was extended to the SSWI fields produced by
HIRLAM-INM during the last month of January. The
graph (top on the right) shows that, as expected,
the skill scores are now somewhat worse. The
thick (thin) dashed lines correspond to r.m.s.e.
(scale on the left y-axis) and to the linear
correlation coefficient (scale on the right
y-axis) for the forecast mode (hindcast
mode). The solid lines correspond to the bias
(scale on the left y-axis). The most remarkable
result is the difference in bias between both
sets of model fields. The thick solid line lays
about 10 W/m2 above the thin solid line. It turns
out that this discrepancy concentrates over the
coastal areas and is less important inland (map
on the right). It is likely that this difference
can be traced back to the different analyses used
in either case, in particular to the humidity
field analyses.