This talk presents a prototype data-driven wildfire spread simulator capable of correcting inaccurate predictions of the fire front position and of subsequently providing an optimized forecast of the wildfire behavior. The potential of the prototype simulator is highlighted on a reduced-scale controlled grassland fire experiment.
The prototype simulator features:
● an Eulerian front-tracking solver that treats the fire as a propagating interface at regional scales
● a series of observations of the fire front position
● a data assimilation algorithm based on an Ensemble Kalman Filter (EnKF), which features a state estimation approach directly correcting the fire front position.
Best Student Paper Award
➞ Rochoux, M.C., Emery, C., Ricci, S., Cuenot, B., and Trouvé, A. (2014) Towards predictive simulation of wildfire spread at regional scale using ensemble-based data assimilation to correct the fire front position, in Fire Safety Science - Proceedings of the Eleventh International Symposium, International Association for Fire Safety Science.
1. Towards predictive simulations of
wildfire spread at regional scale
using ensemble-based data assimilation to correct
the fire front position
M. Rochoux, C. Emery, S. Ricci, B. Cuenot, A. Trouvé
11th International Symposium on Fire Safety Science
U. Canterbury, February 9-14, 2014
3. 3 3INTRODUCTION ● ● ● ●
Regional-scale modeling viewpoint
3
Favone wildfire (30 ha) • Objective: predict the time-evolving position
of the fire front
FOREFIRE
Coupled
Observed
Semi-empirical models: Rothermel [1] , Balbi [2]
[1] Rothermel, Tech. Rep. USDA Forest Service (1972)
[2] Balbi et al. Comb. Flame (2004)
[3] Filippi et al. Proc. Comb. Inst. (2013)
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
FOREFIRE/MESO-NH [3]
Two-way fire/atmosphere coupling
• Surface fire spread model
• Surface/atmosphere fluxes
• Atmospheric transport/chemistry
4. 4INTRODUCTION ● ● ● ●
Scope of the thesis
4
Sources of epistemic uncertainty
• Environmental conditions
• Physical modeling inaccuracies
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
• OBJECTIVE: Quantify and reduce uncertainty in ROS modeling
for wildfire spread forecasting
Data assimilation
Measurements Front-tracking
Environmental conditions
Sparse due to the opacity
of the thermal plume
➔ Optimal combination of measurements and model
Improved forecast of
the fire front position
Weather forecast
Atm. chemistry
Hydrology
Biomechanics
5. ①➀ An eye onto data assimilation
②➁ Strategy for wildfire spread forecast
• Data assimilation components
• Estimation of error statistics
➂ Data assimilation experiments
• Synthetic cases
• Real-case study
➃ Concluding remarks
5Outline 5
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
6. reality
Model forecast
Diagnostic
Measurements
Analysis
Time
Sequential approach
• Estimation of modeling/observation error statistics: distance to reality
• Explicit formulation ➔ Ensemble Kalman filter (EnKF)
➔ Correction of the forecast control variables
= + K [ - ]
Distance to
observationsKalman gain
matrix
Stochastic characterization
• Ensemble of simulations of
the forward model
• Statistical evaluation of
error covariance matrices
➀ DATA ASSIMILATION ●
Kalman filter and non-linear extension
6
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
7. ①➀ An eye onto data assimilation
②➁ Strategy for wildfire spread forecast
• Data assimilation components
• Estimation of error statistics
➂ Data assimilation experiments
• Synthetic cases
• Real-case study
➃ Concluding remarks
7Outline 7
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
8. 8
➁ STRATEGY ● ●
Data assimilation components
8
Observations
Observed
front
x
y
Uncertainty
for each front
marker
Type of observations
Front-tracking simulator
Control variables
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
9. Observations
Model counterpartFIREFLY
Front
distance
• Rothermel ROS model
• Level-set-based solver
Simulated front
Observed
front
(x1
, y1
)
(x2
, y2
)
(x3
,y3
)
(x4
, y4
)
(x1
O
,y1
O
)
(x2
O
,y2
O
)
ROS model parameters
Initial condition
• Wind magnitude/direction
• Fuel moisture content
• Fuel surface/volume
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
9
➁ STRATEGY ● ●
Data assimilation components
9
Type of observations
Front-tracking simulator
Control variables
10. Observations
Model counterpartFIREFLY
Front
distance
Simulated front
Observed
front
(x1
, y1
)
(x2
, y2
)
(x3
,y3
)
(x4
, y4
)
(x1
O
,y1
O
)
(x2
O
,y2
O
)
ROS model parameters
Initial condition
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
10
➁ STRATEGY ● ●
Data assimilation components
10
Type of observations
Front-tracking simulator
Control variables
AnalysisCurrently: model state estimation
• Correction of the fire front location
• Spatial deformations of the fire front
Ensemble Kalman filter
Parameter estimation
Model state
estimation
Previously: parameter estimation
• Sequential approach
• Mean correction along the fireline
11. 11
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
11
➁ STRATEGY ● ●
Estimation of error statistics
11
ISOTROPIC
Uncertainty: ignition location
• Generation of the forecast ensemble
• Representation of the sources of uncertainties
• Description of error correlations along the fireline
Analysis
True
Observation
Forecast
True
Observation
Distance to simulated front marker [m]
Errorcorrelation
Correlations y-y
Rochoux et al., NHESS Special
Issue, part II, in preparation
12. ANISOTROPIC
Uncertainty: wind magnitude and direction, fuel moisture content, fuel surface/volume,
fuel depth, ignition location
12
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
12
➁ STRATEGY ● ●
Estimation of error statistics
12
• Generation of the forecast ensemble
• Representation of the sources of uncertainties
• Description of error correlations along the fireline
Analysis
True
Forecast
True
Observation
Distance to simulated front marker [m]
Errorcorrelation
Correlations y-y
Rochoux et al., NHESS Special
Issue, part II, in preparation
ISSUE: How to accurately estimate error correlations?
13. ①➀ An eye onto data assimilation
②➁ Strategy for wildfire spread forecast
• Data assimilation components
• Estimation of error statistics
➂ Data assimilation experiments
• Synthetic cases
• Real-case study
➃ Concluding remarks
13Outline 13
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
14. 1414
③ EXPERIMENTS ● ●
Synthetic cases
14
• Sensitivity to the location
of observed markers
True
Forecast
Analysis
True
Observation
IDEAL CASE
Improved
performance in the
non-informed
section with data
assimilation
PRACTICAL CASE
Opacity of the fire
thermal plume
Observation
Analysis
True
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
15. 1515
③ EXPERIMENTS ● ●
Synthetic cases
15
• Forecast performance: limited persistence of the initial condition
• Anisotropic case with temporally-varying wind conditions
• Multiple assimilation cycles at 150 s intervals
consequence of
stopping the analysis
analysis
every cycle
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
State estimation needs to
be renewed frequently
in accordance to the
persistence of the initial
condition
16. 1616
③ EXPERIMENTS ● ●
Real-case study
16
• Controlled grassland fire experiment
• Quasi-homogeneous short grass ➔ Moisture content: 22 %
• Mean ROS = 2 cm/s ➔ Max. ROS = 5 cm/s
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
1min32s50s 1min46s1min04s 1min28s
Time
Wind 1m/s
x [m]
y [m]
Extraction of the 600 K
iso-temperature at 14 s
intervals
Estimated error: 5 cm
(1 % burned area)
Front
17. Free runAnalysis
MeasurementForecast
1717
③ EXPERIMENTS ● ●
Real-case study
17
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
1min32s50s 1min46s1min04s 1min28s
Time
Free runAnalysis
Measurement
IC
• 100 members
• Uncertain parameters
• Grass moisture content
• Grass surface/volume
• Wind magnitude
• Wind direction
• Fuel depth in 4 zones
• Ignition location
Free run
Analysis
Measurement
Forecast
18. ①➀ An eye onto data assimilation
②➁ Strategy for wildfire spread forecast
• Data assimilation components
• Estimation of error statistics
➂ Data assimilation experiments
• Synthetic cases
• Real-case study
➃ Concluding remarks
18Outline 18
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
19. 1919
④ CONCLUSION ● ●
Contributions of this work
19
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
Data-driven fire modeling is capable of
• Correcting inaccurate predictions of the fire front position
• Providing optimized forecast of the wildfire behavior
State estimation approach
• Non-uniform correction of the fire front location
• Limited persistence of the initial condition
FORECAST
State
Parameter
State
Parameter
ANALYSIS
Rochoux et al., PROCI,
submitted for publication.
20. 2020
④ CONCLUSION ● ●
Contributions of this work
20
Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
Data-driven fire modeling is a promising novel approach to
• Reduce modeling uncertainties by integrating fire modeling
and fire sensing technologies
• Take advantage in recent progress made in sensor
technology (in-situ/airborne/spaceborne)
Application to real-time emergency response
management
• Monitoring of fire growth
• Forecasting of fire location/size and smoke transport
21. Mélanie ROCHOUX 14 février 2014IAFSS conference, U. Canterbury
Thank you! Any question?
Contact: melanie.rochoux@graduates.centraliens.net
Ph.D. thesis (soon available online)
Towards a more comprehensive monitoring of wildfire spread: Contributions
of model evaluation and data assimilation strategies.
Notes de l'éditeur
Good morning everyone!
Rothermel/Balbi =>modèle ROS intégré dans FOREFIRE qui ensuite propage le front
Favone wildfire (8 July 2009, Corsica).
Blue line : simulation 50 min after ignition
Green line : simulation 4 hours after ignition (coupling mode)
Yellow line : Green line in non-coupled mode (surface mode)
Red line : final observed fire front
Heat transfer to vegetation
Steady-state assumption
Lack of modeling for fire/atmosphere interactions
12 minutes
Matrice de gain = pondération barycentrique de la distance entre obs et prévision du modèle en fonction des statistiques d’erreurs
12 minutes
PARAMETERS (UNIFORM)
➔ High-spatial resolution of environmental conditions not available
➔ High dimensionality of the data assimilation problem
Observation operator
Including FIREFLY model integration
Correlation between errors on parameters and errors on observed quantities
Observation operator => Error correlation between front marker positions