NAQFC: National Air Quality Forecast Capability, Verification

Download Report

Transcript NAQFC: National Air Quality Forecast Capability, Verification

NCEP NAQFC Verification System
Perry Shafran ,Binbin Zhou, Caterina Tassone
Jeff McQueen,Geoff DiMego
NOAA/NWS/NCEP/EMC
Jerry Gorline
NOAA/NWS/MDL
15 June 2015
1
NCEP Verification System
2
NCEP Verification System
- made up of several parts:
Editbufr – edits input prepbufr files to output only the preferred obs to be verified
Prepfits – interpolates model output to ob location and writes out model-ob pairs
Gridtobs – calculates partial sums and writes out sums in VSDB file
FVS (Forecast Verification System) – calculates final statistics and produces plots
3
The NCEP Verification System
4
NAM vs NAM-X Temperature Bias vs Pressure Profile Graphic
NAQFC Verification
Parameter
Model
Observation
Statistic
Web site
1 & 8 hr avg
ozone hrly
CMAQ
AIRNOW
monitors
Partial sum (bias,
RMSE..)
http://www.emc.ncep.noaa.gov/mm
b/aq/fvs/web/html
1 & 8 hr avg
ozone daily
maximum (4z-4z)
CMAQ
AIRNOW
monitors
Threshold skill scores
(fraction correct, CSI)
50,60,65,70,75,85,105,
125,150 ppb
““
1 hr avg PM2.5
hrly
CMAQ
AIRNOW
monitors
Partial sum (bias,
RMSE..)
““
1 & 24 hr avg
PM2.5 daily
maximum
CMAQ
AIRNOW
monitors
Threshold skill scores
(fraction correct, CSI)
12,15,35,55,75 ug/m3
““
1 hr column
integrated Smoke
and dust
HYSPLIT
NESDIS GASP
gridded
smoke/dust
mask
Threshold skill scores
(1,2,5,10,15,20 ug/m3)
http://www.emc.ncep.noaa.gov/mm
b/aq/hysplit/fvs/web/html
5
NCEP Verification Regions
6
Editbufr
Input – A standard prepbufr file
What it does – Inputs observations and trims the prepbufr file based on:
1) Ob type – surface obs, raobs, profilers, ships, pibals, satellite, etc
2) Time window – center obs +/- 15 min, 30 min, up to 90 min
3) Grid region – Use standard NCEP grid number or user-defined grid
box
Output – A thinned prepbufr file
7
Prepfits
Input – the thinned prepbufr file from Editbufr
- all the model grib files that verify at this time
What it does: 1) Reads in model variables
2) Reads in obs variables
3) Interpolates model variables to the ob site using a bilinear
interpolation
4) Writes out the model and obs pair at the same location to a
new bufr file called the “prepfits” file, for each forecast time
Output – The prepfits file – the model-ob pairs file
8
Gridtobs
Input: The prepfits file
What it does: 1) Reads in model-ob pairs
2) Reads in user-defined control file that controls:
a) What variables to verify (Z, T, etc...)
b) What ob types to verify (surface, upper air, etc...)
c) What forecast hours to verify (12-hr, 24-hr, etc...)
d) What levels to verify (1000 mb, 500 mb, etc...)
e) What statistic type is verified (RMSE/bias types,
FHO, etc...)
f) If FHO – what thresholds to verify
3) Calculates and writes out the partial sums
An ensemble version (gridtobsE) is available as well.
Output: A VSDB (Verification Statistical Data Base) file
9
Available variables
Upper air (raobs, profiles, aircraft)
- geopotential height
- temperature
- specific humidity (raobs only)
- relative humidity (raobs only)
- wind (vector, speed, direction)
- PBL height (discussed later)
- CAPE, CIN, LI
- Precipitable water
- Tropopause height
Surface/near surface
- sea level pressure
- 2-m temperature
- 2-m dew point
- 2-m relative humidity
- 10-m wind
- visibility
- heat index
- wind chill
- total cloud
- Surface Ozone
- Surface PM2.5
- Total Column Smoke and Dust
10
FVS – Forecast Verification System
Input – VSDB files
What it does: 1) Reads in VSDB files
2) Reads in user-defined control file
3) Reads in menu-based options to control
a) The statistic (RMSE, bias, etc...)
b) The look of the plot (axis labels, plot title, etc...)
c) Plot smoothing
d) Output type (plot on screen, postscript, gif, etc..)
4) Calculates final statistic
5) Produces plot using on Gempak-based subroutines
Output: A statistical plot!
11
FVS – Forecast Verification System
Statistics available
- Root mean square error
- Simple ob vs. model plot
- Bias
- Anomaly correlation
- Variance
- Standard deviation
- Difference plot
- Equitable threat score
- Bias score
- Mean absolute error
- Mean squared error
- S1 score (g2g only)
Plot types
- Time series
- Diurnal cycle
- Profiles
- Histograms
- ROC diagrams
- Reliability diagrams
- Talagrand diagrams
- Equally likely diagrams
- Ensemble spread
12
Daily Ozone and PM Maximum
Threat Score Plots
July 2014
CMAQ-Prod : CB-IV chemistry, no Aerosol
CMAQ-V4.6.2: CB05 chemistry, AERO-IV
CMAQ-V4.6.3: CB05 chemistry, AERO-IV, real-time smoke/dust sources
13
Ozone, Aerosol and PBL Verification System at NCEP
Observations
Model output
RAOB
Aircraft
TKE PBL
Mix Layer Ht
PBL calculation
PBL height
Profiler
AIRNOW
NAM
Ri PBL/no fluxes
Ri PBL/fluxes
PM2.5
Aerosol
Ozone
Ozone
CMAQ
Forecast Verification System
Statistics
Model PBL height
NAM/NAMB:
1. Based on TKE profile (TKE=0.2)
2. Based on Ri number approach (critical Ri=0.25); no fluxes (same as from obs)
3. Mixed layer depth (where TKE production equals dissipation)
CMAQ:
ACM2 PBL – enhanced Ri number approach; sfc fluxes used
14
PBL Analysis : Observations
 RAOBS, ACARS and CAP Profilers:
Bulk Richardson Number (RIB) approach (Vogelezang and Holtslag,1996):
PBL height is defined as the level where the RIB exceeds the Richardson
Critical Number (0.25)
RIB 
(g/ θvs)(θvh  θvs)(h  z )
s
(u h u s)  (vh  vs)  bu2
2
2
ACARS
Derivation of PBL height from
aircraft observations (M.Tsidulko)
Evaluation of ACARS
- ACARS level data (U,V,T,P)
- Surface Observation
- Moisture analysis from model
- QC
- Evaluation with DC 2009 experiment:
Good agreement with measurements
(Radiosondes, lidars). Underestimates PBL
height in early afternoon and some QC control
issues
- Good diurnal variation of PBL height
15
• ACARS – hourly: diurnal cycle,
• 12 km TKE PBL higher in NAMB for CONUS, EAST, WEST
• 4 km TKE PBL lower than 12 km PBL for CONUS and EAST; less evident for WEST
• Almost no difference for RI PBL
16
Improvements to ACARS PBLH Verification
The data used for the Richardson number computations do not represent a truly vertical
profile
1. Distance between PBL Height and surface measurement location, especially for
inhomogeneous surface (e.g. sea-land transition)
2.Vertical profiles of u,v,T and q different due to
difference in aircraft trajectories.
Miami-PBLH within 9 minutes
20.51z
PBLH=500m
21.00z
PBLH=1350m
Attach closest surface observation to aircraft track --->
more representative of underlying surface influence
Flag observation when PBLH over water --->
for increased weighting of background
17
Cooperative Agency Profilers (CAP)
Boundary Layer profiler and RASS data acquired in near-realtime.
BL profilers: Doppler radars that measure vertical profiles of horizontal winds.
Wind measurements from 100m up to 3-5 km; range from 60 to 400m.
RASS (Radio Acoustic Sounding System): virtual temperature between 200 m and 1 km
Derivation of PBLH from CAP
 Attach a surface
 Interpolate temperature to winds levels and wind to temperature levels
 Use surface pressure and hydrostatic equation to derive pressure at data levels
 Compute PBLH using Richardson Number method
 QC: RIN > 0.25 between 2nd and last level and z(2)-z(1) < 300m (11 stations left)
Daily average of CAP Profiles 29 Oct - 5 Nov 2010
30
Evaluation: (29 Oct to 5 Nov 2010)
- Statistics  no diurnal variation
detected because of lack of lower
and higher levels
- Comparison with nearby ACARS
- Comparison with BL profilers
25
20
15
10
5
18
0
0
1
2
3
4
5
6
7
8
9
10
avalaible CAP profiles
11
12
13
14
15
16
17
18
n. of PBLH derived from CAP profiles
19
20
21
22
23
Comparison with ACARS
Different behavior for different stations:
Baltimore: CAP always underestimates
PBLH. More studies are needed.
Lunenburg: not enough data
Seattle: big spread
Los Angeles and Sacramento : good
Los Angeles
 Many CAP data have been eliminated by
QC
 High number of ACARS data
 Location of CAP profilers and airport
very close
19
Comparison with Boundary Layer profilers
Boundary Layer Profilers (L.Bianco and Wilczak, 2002)
BL depth is automatically determined using SNR (Signal-to-noise Ratio), vertical velocity
and turbulence intensity.
Good agreement between ACARS (KSMF), CAP and BL profilers at SAC (Sacramento)
and between CAP and BL profilers at CCL (Chowchilla) and at CCO (Chico)
20
Derived PBL Height from ACARS
Time Series Plot
21
NCEP Grid-to-Grid Verification System
22
Prepg2g (shell progra)
Automatically search model & truth data according to path setup
Climate GRIB data file
Region and North America definition files
User-defined Control (created by user)
9 fields, such as:
verification time, forecast hours
obs type
stats types (SL1L2, VL1L2, FHO, etc)
variables (defined by GRIB PDS-5,6,7) , thresholds
Levels
….
VSDB files
Output files store partial sum of statistics over region (s)
23
Example: SL1L2 VSDB Record (total cloud ~ AFWA satellite data)
Hdr
V01
V01
V01
V01
mod
NAM
NAM
NAM
NAM
fcst_hr valid_time truth grid stats variable level
06 2010112100 AFWA G212 SL1L2 TOTCLD ATMOS =
12 2010112100 AFWA G212 SL1L2 TOTCLD ATMOS =
18 2010112100 AFWA G212 SL1L2 TOTCLD ATMOS =
24 2010112100 AFWA G212 SL1L2 TOTCLD ATMOS =
grid#
19931
19931
19931
19931
f
58.9
59.3
60.1
60.8
o
40.5
40.5
40.5
40.5
fo
2831.6
2834.7
2845.5
2894.4
ff
5349.3
5383.9
5476.7
5545.9
oo
2691.6
2695.6
2691.6
2691.6
f,o – forecast and observation average over all grid# . Based on f,o,f*o,f*f and o*o,
SL1L2 stats (rmse, bias, etc can be computed)
e.g.: Bias = f – o ; rmes = sqrt (ff -2*fo + oo);
Example: FHO VSDB Record (total cloud ~ AFWA satellite data)
Hdr mod
V01 NAM
V01 NAM
V01 NAM
V01 NAM
V01 NAM
V01 NAM
V01 NAM
V01 NAM
V01 NAM
fcst-hr valid-time truth
12 2010112100 AFWA
12 2010112100 AFWA
12 2010112100 AFWA
12 2010112100 AFWA
12 2010112100 AFWA
12 2010112100 AFWA
12 2010112100 AFWA
12 2010112100 AFWA
12 2010112100 AFWA
grid
G212
G212
G212
G212
G212
G212
G212
G212
G212
stats>thres
FHO>10
FHO>20
FHO>30
FHO>40
FHO>50
FHO>60
FHO>70
FHO>80
FHO>90
variable
TOTCLD
TOTCLD
TOTCLD
TOTCLD
TOTCLD
TOTCLD
TOTCLD
TOTCLD
TOTCLD
level
ATMOS =
ATMOS =
ATMOS =
ATMOS =
ATMOS =
ATMOS =
ATMOS =
ATMOS =
ATMOS =
grid#
19931.
19931.
19931.
19931.
19931.
19931.
19931.
19931.
19931.
F
.671
.643
.624
.606
.587
.560
.535
.506
.473
H
O
.549 .758
.459 .633
.390 .533
.328 .444
.277 .368
.223 .298
.178 .237
.128 .171
.086 .109
F, H, O – forecast, hit and observed rates ( x grid# = forecast, hit and observed grid
points), based on which POD, FAR, CSI, ETS can be computed
e.g.: POD = H/O; FAR=1-H/F; CSI=H/(O+F-H)
24
G2G Verifications at EMC/NCEP
Model/System
NCEP-GFS, ECWMF and other
GFS (UKmet, Canada, France)
NAM, GFS
Analysis as truth
Variables
GDAS
T, RH, Wind, etc at 850mb, 500mb
and 250mb
RTMA
Sfc T, Td, Wind, SLP, etc
NAM, GFS
AFWA satellite data
Total cloud
NAM, GFS
CLAVR satellite data
Total cloud
NAM, RUC, High-res WRF
MOSAIC radar data
NAM, RUC
ADDS
SREF, VSREF
ADDS
SREF, VSREF
MOSAIC radar data
HYSPLIT
NESDIS smoke/dust detection
CMAQ
NESDIS AOD detection
SCIPUFF
MM5-dose analysis
Reflectivity & echo-top
Visibility/fog
Visibility/Fog probability
Reflectivity probability
smoke/dust AOD
Aerosol Optical Depth
SCIPUFF dose probability
25
26
27
28
NWS MDL Verification
8h avg Daily Maximum Ozone Spatial Plots
Four-day
outbreak of
2008, day 3,
188 observed
exceedances
FC=0.773
TS=0.362
POD=0.766
FAR=0.593
Predicted in dark blue
29
Observed as red dots
NWS MDL Verification
1h avg daily Max PM Spatial Comparisons
30
NWS/MDL AQ Verification
2x2 contingency definitions
Predicted




Observed
Yes
Yes
a
yy
No
c
ny
No
b
yn
d
nn
FC = (a + d)/(a + b + c + d)
TS = a /(a + b + c)
Thresholds Used:
POD = a/(a + c)
Ozone: 76 ppb
FAR = b/(a + b)
Aerosols: 35 ug/m3
31
NWS MDL Verification
Long term NAQFC PM Bias Trends
32
NWS/MDL Verification
CMAQ PM2.5 scatterplots
CMAQ Raw PM2.5 predictions
Feb. 2015
CMAQ Bias-Corrected PM2.5 predictions
Feb. 2015
33