quantifying the uncertainty in volcanic ash forecasts · quantifying the uncertainty in volcanic...
TRANSCRIPT
Copyright University of Reading
QUANTIFYING THE UNCERTAINTY IN VOLCANIC ASH FORECASTS
1
Department of Meteorology
Helen Dacre1, Natalie Harvey1, Kelsey Mulder1, Nathan Huntley2,
David Thomson3, Helen Webster3, Peter Webley4, Don Morton4
1. University of Reading, 2. University of Durham, 3. Met Office, 4. University
of Alaska, Fairbanks
USERS INFER UNCERTAINTY IN
VOLCANIC ASH FORECASTS
2
Area of no-fly zone
Heat map showing
no-fly zones
Mulder et al. 2017 (WCAS)
Q. Given the forecast below, draw a no-fly zone
VOLCANIC ASH PREDICTION • Real time forecasts and future risk to infrastructure
• Airborne and surface concentrations
3 Meteorology Volcano characteristics Atmospheric physics
Dacre et al. 2013 (ACP)
QUANTIFYING UNCERTAINTY
Ensemble meteorology
Ensemble model
parameters
Ensemble dispersion
models
4
1. ENSEMBLE METEOROLOGY
5
METEOROLOGICAL UNCERTAINTY
6 Ash column load forecasts for two forecasts (top)
12 UTC on 7 May, (bottom) 00UTC on 8 May 2010
Ensemble member # 1 Ensemble member # 2
bifurcation
Satellite ash cloud
Missing
ash
TRAJECTORY SPREAD
• Particle trajectories rapidly diverge after the trajectories
encounter a bifurcation point 7
Flow separation along 20 particle trajectories released at 06UTC on 6 May
2010 transported using different 72 hour ensemble forecasts
FLOW SEPARATION
• In some meteorological situations the trajectories diverge
while in other situations they remain close together 8
Low wind-speed at
source
Evolution of ensemble spread for 18 simulations initialised between 15 April
and 7 May. Colours show along-trajectory accumulated flow separation
2. ENSEMBLE MODEL PARAMETERS
9
EXPERT ELICITATION
10
18 parameters
identified, 6
ESP’s & 12
internal
parameters
15 parameters
varied
independently
1700 parameter
sets created
Harvey et al. 2017 (NHESS)
EMULATION
11
• An emulator is a simple approximation of the complicated
model that can be evaluated almost instantly
• Rather than approximating the entire output, concentrate on
average column integrated mass in given regions
• 3-4 regions / hour results in 75 emulators
Region 1
Region 2
Region 3
SEVIRI satellite retrieved ash column
loading 0 UTC 14 May 2010
EMULATION
12
• Emulator estimates the expected value and the variance
for the summary f(x) should we run NAME at parameter
choice x
One dimensional example of an emulator. The points represent 6 NAME
simulations of ash column loading at parameter choices x. Emulator
prediction (black) ±3 standard deviations (red)
Average column
integrated ash
loading in region, f(x)
Parameter choice, x
IDENTIFY ACTIVE PARAMETERS
13
• Emulators improved by focussing on important parameters
• Parameters removed in turn and R2 calculated, parameters
whose removal cause the smallest change were eliminated
• Only a small number of parameters contribute to the
uncertainty in each region
Parameter
Number of regions
where parameter is
active
Plume height 75
Mass eruption rate 75
Standard deviation of velocity for free tropospheric turbulence 61
Precipitation rate required for wet deposition 58
Particle size distribution scale parameter 18
Lagrangian timescale for free tropospheric turbulence 15
All other parameters 0-4
3. ENSEMBLE DISPERSION MODELS
14
ENSEMBLE DISPERSION MODELS
15 Animation of volcanic ash simulations using 3 dispersion models. No ash in
any simulation (dots), ash in all simulations (hatching)
SUMMARY METEOROLOGICAL UNCERTAINTY
• Nearby ash particle trajectories can rapidly diverge leading poor
forecast accuracy for deterministic forecasts which do not represent
variability in wind fields at the synoptic-scale
INPUT AND INTERNAL PARAMETER UNCERTAINTY
• Statistical approximation (emulator) of NAME used to identified which
parameters contribute the most to prediction uncertainty allowing us to
focus areas for improved observations or model development
MULTI-MODEL UNCERTAINTY
• Using consistent meteorology and ESPs allows us to visualise multi-
model uncertainty and communicate forecast confidence where models
agree
16
EXTRA SLIDES
17
VOLCANIC ASH HAZARD • More than 80 volcanoes in Europe with over 1200 recorded eruptions
• Relatively small eruptions can cause major disruption
• Icelandic volcanoes erupt ~ every 5 years
18
RISK APPETITE
19
INPUT AND INTERNAL
PARAMETER UNCERTAINTY
20
• Multi-parameter sensitivity analysis has 2 main advantages
• Amount of parameter space sampled is larger
• Includes interactions between parameters
• Disadvantage is that dispersion models run too slowly to
evaluate many (>3) parameter choices at the same time
2 parameters
= 102
3 parameters
= 103
1 parameter
= 101
4 parameters
= 104
VALIDATE EMULATORS
21
• Able to perform an additional 1500 Fast NAME simulations
• Over the 75 regions, the proportion of successful
predictions from the validation ranged from 94.5% to 99%
Leave-one-out validation plot of emulator for 1st output (region 1).
Emulator expected value for parameter sets x_i (black) ± 3 standard
deviations (blue), NAME output at each parameter set (red).
Region 1