1. Simulation of the 2009 Harmanli fire (Bulgaria) Nina Dobrinkova Georgi Joradanov, Jonathan Beezley, Adam Kochanski, Jan Mandel and Bedrich Sousedik LSSC 6-10 June, Sozopol, 2011

2. In this presentation … 1)Introduction 2)WRF- fire model, some basics 3)Simulations done until now 4)Conclusions LSSC 6-10 June, Sozopol, 2011

3. Forest fire statistics - Europe The number of fires since 1980 according to statistics done for the southern member states has increased rapidly in the last 20-25 years LSSC 6-10 June, Sozopol, 2011

4. Forest fire statistics - Bulgaria Bulgaria’s statistic about forest fires 1994 to 2006 1971 to 2006 considerable increase of the number of fires after 1990 (more than 1000 in year 2000) more than 30 times increase of the burned area in the recent years LSSC 6-10 June, Sozopol, 2011

5. WRF-Fire basics (1) Mathematically the fire model is posed in the horizontal (x,y) plane. The fire propagation is in semi-empirical approach and it is assumed that the fire spreads in the direction normal to the fireline. This is given from the the modified Rothermel’s formula: S=min{B 0,R 0 + ɸ W + ɸ S }, where B 0 is the spread rate against the wind; R 0 is the spread rate in the absence of wind; ɸ W is the wind correction ɸ S is the terrain correction LSSC 6-10 June, Sozopol, 2011

6. WRF-Fire basics (2) Once the fuel is ignited, the amount of the fuel at location (x, y) is given by: Where : t is the time; t i is the ignition time; F 0 is the initial amount of fuel; T is the time for the fuel to burn down to 1/e of the original quantity LSSC 6-10 June, Sozopol, 2011

7. WRF-Fire basics (3) From slides (1) and (2) we have idea about the plane, where the fire will spread and the fuel which we want to ignited, but we also need the heat flux, which is inserted as the time derivative of the temperature, while the latent heat flux as the time derivative of water vapor concentration. This scheme is required because atmospheric models with explicit time stepping, such as WRF, do not support flux boundary conditions. LSSC 6-10 June, Sozopol, 2011

8. WRF-Fire basics (4) From the previous three slides we have the plane of the fire, the ignited fuel, the heat flux, but we also will need the burning region at time t. It is represented by level set function ɸ, as the set of all points (x, y) where ɸ (x, y, t) < 0. The level set function satisfies a partial differential equation for dynamic implicit surfaces: Where is the Eucledian norm of the gradient of ɸ. LSSC 6-10 June, Sozopol, 2011

9. Simulation results (1) We used for our first simulation we used WRF-Fire v.3.2. We did a domain of size 4 by 4 km, with horizontal resolution of 50 m, for the atmosphere mesh, 80 by 80 grid cells and with 41 vertical levels from ground surface to 100hPa. We didn’t use nesting. That domain was located 4 km west from village Zheleznitsa in the south-east part of Sofia district. The domain was covering the lower part of the forest of Vitosha mountain. The ignition line was in the center of the domain 345 m long and the ignition starts 2 seconds after the simulation biginning. LSSC 6-10 June, Sozopol, 2011

10. MN&A 20-24 August, Borovec, 2010

11. Simulation results (2) The first simulation was done for ideal case in the observed area, but we wanted to test WRF-Fire with data about real fire. We set two domains the first was covering area of 48 km 2 with resolution 300m (160x160). This domain was producing boundary and initial meteorological conditions for the inner domain and in this domain there were no fire simulations. The inner domain was located in the middle of the coarse domain. The resolution in D2 we set as 60m and the area covered is 9.6 km 2 (161x161). D2 was centred on the fire ignition line and it was covering the areas of villages Ivanovo, Leshnikovo and Cherna Mogila. This area was located in South-East Bulgaria close to the Bulgarian-Greece border.This particular area was chosen because on 14 th till 17 th of August 2009 there was a wildland fire spreading in wide area and burning everything in the range. We manage to run this simulation only for 15 minutes. LSSC 6-10 June, Sozopol, 2011

12. The fire ignition is set for the date of 14.08.2009 The ignition start 60sec after the simulation start. The ignition line is 1.3km long and 200m wide. Simulation results (2.1) LSSC 6-10 June, Sozopol, 2011

14. ICFFR 15-18 November, Coimbra, 2010

15. LSSC 6-10 June, Sozopol, 2011

18. Simulation results (3) The objective of this third simulation is to present the simulation capabilities of WRF-Fire model with real input data. Atmospheric model was run on 2 domains with 250m and 50m resolution 41 vertical levels were used The fire module, coupled with the atmosperic domain is run on 5m resolution with 0.3s time step Simulated burned area and actual data from the Ministry of Agriculture, Forest and Food showed good comparison LSSC 6-10 June, Sozopol, 2011

19. Simulation results (3.1)

20. Simulation results (3.2) the heat flux (red is high) burned area (black) atmospheric flow (purple is over 10m/s) Note the updraft caused by the fire Ground image from Google Earth LSSC 6-10 June, Sozopol, 2011

27. Data sources (1) With all input data available, we use the WRF preprocessor (WPS) to produce the model input Meteo input - global reanalysis data from the U.S. NCEP database (Grib format) Static data fields describing the surface properties – most of this data is from the model global dataset with resolution of arr.1km Topography data is very important and much more detailed source is the (90m resolution) for the area of Harmanli is used from the Shuttle Radar Topography Mission (SRTM) - the data is GIS raster format (DTED) LSSC 6-10 June, Sozopol, 2011

28. Data sources (2) Available fuel data - created using GIS and data from the Corine2006 landcover project (for whole Bulgaria) - the data set has 100m resolution which makes it suitable for fire behavior simulations - we assign each area a fuel category using the 13categories standard Anderson fuel models LSSC 6-10 June, Sozopol, 2011

29. Parallel performance Computations were performed on the Janus cluster at the University of Colorado. The computer consists of nodes with dual Intel X5660 processors (total 12 cores per node), connected by QDR InfiniBand The model runs as fast as real time on 120 cores and it is twice faster on 360 (real time coef. = 0.99) LSSC 6-10 June, Sozopol, 2011 Cores 612243660120240360480720 Real time coefficients 10.599.213.912.751.640.990.610.440.370.31

30. Conclusion (1) The simulation of the fire near by Harmanli used: Fuel data from CORINE satellite Very fine, 50m resolution: Need 120 cores to be as fast as real time, 360 cores to be twice faster: LSSC 6-10 June, Sozopol, 2011

31. Conclusion (2) We have demonstrated wildfire simulation based on real data in Bulgraria from satellite measurement and existing GIS databases The simulation provided a reasonable reproduction of the fire spread The simulation showed correct fire line propagation, and it can give forecast and valuable information for future firefighting actions in different areas with different meteorological conditions The model can perform faster than real time at the required resolution, thus satisfying one basic requirement for a future prediction usage LSSC 6-10 June, Sozopol, 2011

32. Nina Dobrinkova Institute of Information and Comunication Technologies - Bulgarian Academy of Sciences, nido@math.bas.bg Thank you for your attention! LSSC 6-10 June, Sozopol, 2011