Anda di halaman 1dari 31

Forecasting Wind-Driven Wildfires Using An Inverse Modelling Approach

Oriol Rios, Wolfram Jahn, Guillermo Rein

Numerical Wildfires Workshop Cargse, 16-5-2013

Outline

Background idea Methodology Forward model Optimization. Tangent lineal model & automatic differentiation Synthetic validation Cases explored Fire fronts Wind speed Wind speed and direction Fuel depth Perturbed data Conclusions & Further Work

Background Idea

Hard to gather information to initialize models in operative situations Complex model require high computational capacity and time Wildfire responders need forecasting tools M. Rochoux et al., J. Mandel et al. started using data assimilation in wildfires

La Jonquera, 2011

La Riba, 2011

Background Idea

Jahn et al. successfully used DA to forecast fire in compartments The algorithm is based on the fact that invariants exists for a certain amount of time ex: Entrainment coefficient

W. Jahn et al. Forecasting fire dynamics using inverse computational fluid dynamics and tangent linearisation, Advances in Engineering Software

Background Idea

Invariants exists and represent one or more physical quantities (i.e. wind speed or fuel properties). Use a simple yet reliable model to explore DA capacities for winddriven wild fires. Versatile DA algorithm regarding available data (invariants reversibility) Ensure positive lead time
Input fire fronts positions (airborne, satellite) during an assimilation window to identify the invariants Data min. Invariants

Forward model

Airborne and Satellite imaged

Source of data

CSIRO-UPC, 2008 FuSE project - Bushfire CRC Ngarkat CP experimental burnings (SAus)

Pliades SAT La Jonquera, 2011 MODIS/Google

The forward model: Rothermels+Huygens

Rate of spread (surface fire)

parameters

11 variables (7+4)

variables

The forward model: Rothermels+Huygens

Huygens principle. Firelets expansion

+ Anderson length-to-breadth correlation 8


G D Richards. The properties of elliptical wildfire growth for time dependent fuel and meteorological conditions. Combustion science and technology,1993

The forward model: Rothermels+Huygens

Generates no trivial perimeters (fuel heterogeneity)

Cost function
Distance between angular correspondent vertexes

10

Optimization

Tangent linear model

First Guess

How to calculate the gradient?


dProgram Automatic differentiation (forward or adjoint) Program aProgram

11

Optimization. Automatic differentiation


sin(x1/x2)

12

Algorithm

13

Validation

I used synthetically data generated with Rothermels+Huygens model (without casting invariants) and initialized with parameters from Behave (Anderson) We studied 4 different invariants cast
4 invariants 3 invariants 3 invariants + wind speed + wind direction + fuel depth

14

3 invariants

Casting the invariants. 1st cast. (4 invariants)

Moisture-fuel Invariant

Wind speed invariant


Wind factor invariant

Wind direction

15

Casting the invariants. 1st cast. (4 invariants)

Step-to-step example

What if the invariants evolution is known?

16

Casting the invariants. 1st cast. (4 invariants)

Convergence to true value

17

Influence initial guess Divergence correction

2nd cast of invariants: 3 invariants & wind speed

Moisture-fuel Invariant

Wind factor invariant

Input Data

Wind direction

19

2nd cast of invariants: 3 invariants & wind speed

20

3rd cast of invariants: 3 invariants & wind (U,)

Moisture-fuel Invariant

Wind factor invariant

Wind direction

Input Data

21

3rd cast of invariants: 3 invariants & wind (U,)

22

4th cast of invariants: 3 invariants & fuel depth (x,y)

RoS linear to fuel depth

Input Data

Wind direction

Length to breadth ratio (Anderson)

I LB
23

4th cast of invariants: 3 invariants & fuel depth (x,y)

24

Data with noise (4th case)

25

Computing time

26

30 min forecast

Positive lead time

Windows of validity

Conclusions

Formulation of the problem is general enough that is suitable to work with many observation (& data contexts). Solution method is fast and positive lead times are already possible with desktop computer. Invariants can be turned into input data for increased accuracy and speed if reliable data arrives The proper invariant cast must be done according to the available data, otherwise multiplicity might be a problem.

27

Further work

Challenge the algorithm with real data (cases needed) Increase the number of invariants to several dozen by means of adjoint modeling approach Assimilate more input data (fire intensity, flame height...)

Move to more powerful optimization routine that require High Performance Computing (eg, evolutionary algorithms) Used more sophisticated forward models (i.e WFDS, FireFoam, ForeFire...)

28

Thank you!
Villemard 1910, National Library of France

Forecast made in 1900 of the fire-fighting in the year 2000.

29

Invariants range
Monte Carlo analysis varying 6 Rothermels variables (20000 sets) within the range established by Scott and Burgan 2005.

30

Invariants influence

Cost function variation when the invariants are perturbed 20% of its base value

31

Invariants influence
Base value and varying range for Rothermels variables

32

Anda mungkin juga menyukai