Numerical Forecasting of Fire Dynamics (Plenary YIC ECCOMAS)

Plenary keynote
Numerical forecasting of fire dynamics: Tomorrow's infrastructure protection

Dr Guillermo Rein School of Engineering University of Edinburgh
Based on the PhD thesis of W Jahn
Seeing the Future
The Tides of Aveiro, Portugal
This is the 48 h tide forecast for the 26th and 27th of March 2012 in Aveiro produced on the 25th This accurate forecasts allow local finishing vessels to program their route and operations ahead of time.
2011 Icelandic ash cloud
May 22nd 2011 forecast made on May 21st
Forecasting Fire Dynamics
Paradigm shift in Emergency Response

Knowledge of future fire conditions spread, smoke, structural collapse Additional layer of essential information to the Fire Service currently non existing Technology for Smart Buildings First adopters of technology expected in critical infrastructure and high profile buildings Predictions need to arrive faster than the event develops (lead time>0)
Lead time: period after the forecast when it is still accurate and valid
Fire Test at BRE commissioned by Arup 2009 4x4x2.4m small premise in shopping mall
190s could this be forecasted ahead of time?
Coupled Fire Mechanisms

SOLID PHASE - fuel Heat conduction Pyrolysis/degradation Flame Spread GAS PHASE flame/smoke Turbulence Combustion Radiation Buoyancy
Fire dynamics are governed by coupled non-linear processes.

NOTE: Combustion is only one of the many important mechanisms in fire dynamics (typical misconception)
Firepower Fuel
Heat release rate (HRR) is the power of the fire (energy release per unit time)
& & & Q (t ) = hc m(t ) = hc m A(t )

WPI
& Q Heat Release Rate (kW) - evolves with time hc Heat of combustion (kJ/kg-fuel) ~ fuel property & m Burning rate (kg/s) - evolves with time & m Burning rate per unit area (m2) ~ fuel property A
Burning area (m2) - evolves with time
Burning rate (per unit area)
from Quintiere, Principles of Fire Behaviour
& q & m = hp
Heat of Combustion
from Introduction to fire Dynamics, Drysdale, Wiley
Lower order modelling: Two-Zone Model

(after Zukoski, 1978)
1D in space and transient in time. Simple model formulated on the fact that the volume of a fire compartment is naturally split in two layers, hot smoke up and cold air down. Smoke layers starts t the ceiling level and descends towards the floor. Mass transfer between both layers is by the fire plume. Mass balance upper layer
Upper layer Plume Lower layer

Leak Firepower growth
Mass balance lower layer
Flow rate of smoke in plume
Jahn et al., Fire Safety Journal, 2011
Higher order model: Computational Fluid Dynamics

3D in space and transient in time State of the art is FDS - Fire Dynamics Simulator:
LES code Mixture fraction Radiation Solid heating
Open source, freely available Developed by NIST (USA) Computational time of a ~10 min fire in a typical single office compartment takes in the order of weeks to solve in a modern desktop PC. The most successful fire CFD code currently in use
Impossible and HPC

Two most common responses we got from experts when we first started to research the topic: 1. That is Impossible, you are wasting your time (note to Young Researchers: this reaction
indicated you are doing something right)
2. No need to research, just take the best CFD model in town and run it as fast as possible using parallel computing, grid and HPC
Round-Robin of Fire Modelling

How accurate is the state-of-the-art? International pool of experts provided a priori (blind) predictions of a largescale fire experiment, 2006 Dalmarnock
Rein et al., Fire Safety Journal 44, 2009
High Density Instrumentation

Deflection Gauges 8 Lasers
20 Heat Flux Gauges
10 Smoke Detectors
10 CCTV
ENLARGE
ENLARGE
ENLARGE
ENLARGE
14 Velocity Probes CCTV
270 Thermocouple
Abecassis-Empis et al., Experimental Thermal and Fluid Science 32, 2008.
Before/After
Average Compartment Temperature
Abecassis-Empis et al., Experimental Thermal and Fluid Science 32, 2008
Predictions of Firepower
Predictions of smoke layer temperature
Predictions of smoke layer position
Round-Robin Lessons
The state-of-the-art of fire modelling is neither accurate nor fast enough for forecasting Brute force forecasting provides excessive uncertainty Firepower growth Q is an essential variable, all others derive from it
Data Assimilation
used in weather forecast
MODEL
SENSORS
FUSION
ANALYSIS Forecasts (with positive lead time)
Forecast of average temperature Current lead time is 3 days 10 years ago was 2 days
Cowlard et al., Fire Technology, 2011
Inverse Modelling
Steps in Data Assimilation: Identify the governing parameters the invariants Quantify via sensor data the value of the invariants Run forecast with those parameters
Weather: quantify initial conditions (IVP) Climate: quantify boundary conditions (BVP) Fire: none of the above The source term, firepower, drives fire dynamics. Q must be estimate first
Inverse modelling is like imitating Sherlock Holmes
Data Assimilation Concept
Invariant values valid
Firepower
New invariant values
Sudden change of conditions (eg, window breakage)
sensor model
Time
Cowlard et al., Fire Technology, 2011
Source term - Invariants

It is a common observation that fires grow as t2 (radial spread at ~ constant rate)
growth parameter ~ constant
& (t ) = h m S 2t 2 = t 2 Q c &
heat of combustion burning rate spread rate
Invariant () is the governing unknown of the problem Other invariants related to ventilation, smoke, etc can be added as well
Inverse Modelling
Forward fire model is A cost function is minimized:
which measures the distance between the observation yi and the output of the forward model yi().
*= argmin(J()) = the invariant values sought
Minimization Technique
For a handful of invariants, gradient techniques are much faster than heuristics Non-linear (NL) gradient technique typically needs >100 iterations Because each iteration requires to run the fire model at least once NL is not as practical for forecasting Solution: linearize Finding: yi() tends to be reasonably linear in compartment fires
Tangent Linear Model (TLM)

Linearizing around guess 0 and replacing into cost function, gradient becomes
Leads to a linear system
Zone Model - Typical Fire Scenario

Compartment 4 x 5 x 2.5 m Mattress fire Sensors: Temperature, ~uniform grid 10 per m3 Synthetic data generated by CFD model (FDS) Invariants: spread rate, entrainment & transport time
This is not the most realistic scenario, specially the high density sensor array. But first attempts ought to be conducted on simple cases before moving towards complex cases
Medium fire (~mattress fire)

Assimilate 5 Data points
Temperature (C)
9 Data points
13 Data points
Upper layer Plume Lower layer

Leak
HRR (kW)
Layer Height (m)
Medium fire (~mattress fire)
Lead time (s)
Slow fire
(ie, large wood slab)
Medium fire
(ie, large mattress)
Lead time (s)
Fast fire
(ie, large polyurethane foam slab)
Lead Time defined as forecast <10% error in upper layer temperature
Lead time (s)
CFD forecast
Same compartment (4 x 5 x 2.5 m, mattress fire) Forward model is LES code FDSv5 Invariants: spread rate, fuel flow (=burning rate) and soot yield Sensors: Temperature and smoke at ceiling height Synthetic data by fine grid FDS (5 cm) Speed up by coarse grid (25 cm)
NOTE: Course grids cannot resolve turbulence and other flow process of importance. But forecasting ought to find a compromise between speed and accuracy. Note that in many weather simulations, for example, Scotland is one single grid cell
Jahn et al., Adv Software Eng, 2012
CFD forecast
TLM vs. BLGF
Comparison of Tangent Linear Method (TLM) and the quasiNewton method Broyden FletcherGoldfarbShanno (BLGF) shows superior performance of the TLM, both in accuracy and in computation time for the problem at hand
Two independent fires
Good convergence!
Effect of sensor type - Assimilation iter.

Using ceiling temperature sensors only Adding ceiling smoke sensor
Poor convergence!
Good convergence!
Unknown location and size of fuel source
Near realtime with CFD forecast

Our CFD forecast do not reach positive lead times yet (current is ~ -4.5 min) because CFD is still too slow compared to real time event. Our works has focused on minimizing the number of iterations for convergence of invariant because:
Each iteration involves a number of parallel CFD runs. A typical CFD run in FDS of a 5 min fire in a single office compartment with a modern PC desktop takes weeks to solve with a grid of 5 cm. With a course grid of 25 cm, it takes 10 min.
High Performance Computing techniques can now accelerate further the inverse problem and reach real time CFD forecast (positive lead times)
Next challenge: CFD forecast using sensor data from a real fire
Data from Dalmarnock Fire Test One High density sensor array for temperature Forward model is FDSv5
CFD forecast of real fire
Rack 1 230 cm high near sofa
Rack 1 160 cm high near sofa
Rack 19 230 cm high near window
Rack 19 160 cm high near window
Jahn et al., IAFSS, 2011
Results show local sensor data can be used to forecast the firepower in the whole compartment
Conclusions
Sensor data from temperature and smoke field used to back calculate the fire growth Methodology is general and independent of the forward model Fundamental step towards the development of forecasting technologies Invariants accurately estimated in <1 min of fire time Positive lead times with zone model (~90 s) Near realtime CFD forecast (~-4.5 min) Coarse grids accelerate forecast up to 100 times without loss of accuracy due to the assimilation of sensor data High Performance Computing techniques can now accelerate further the inverse problem and reach real time CFD forecast (positive lead times)
New perspective
"So easy it seemed, Once found, which yet unfounded most would have thought, Impossible!"
John Milton (1608 - 1674), English poet
Thanks!
Villemard, 1910, National Library of France
Jahn et al, Adv Software Eng, 2012 Jahn et al, IAFSS, 2011 Jahn et al, Fire Safety Journal, 2011 Cowlard et al, Fire Technology, 2011
Paleofuture: forecast made in 1900 of the firefighting in the year 2000

Based on the thesis of my PhD student Wolfram Jah Inverse modelling to forecast enclosure fire dynamics University of Edinburgh, 2010. is in open access at http://hdl.handle.net/1842/3418 Research funded by BRE, EU Alban Scholarship and FireGrid
Three Invariants
Different initial guesses
Good convergence!
All three invariants converge. Soot the slowest
Effect of sensor density

9 sensors
Improvement with CFD assimilation window
General Forecasting Method
Flame Spread vs. Angle
downward vertical spread
upward vertical spread
Upward spread up to 20 times faster than downward spread

Test conducted by Aled Beswick BEng 2009

Numerical Forecasting of Fire Dynamics (Plenary YIC ECCOMAS)

Uploaded by

Copyright:

Available Formats

Numerical Forecasting of Fire Dynamics (Plenary YIC ECCOMAS)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Numerical Forecasting of Fire Dynamics (Plenary YIC ECCOMAS)

Uploaded by

Copyright:

Available Formats

Plenary keynote

Numerical forecasting of fire dynamics: Tomorrow's infrastructure protection

Seeing the Future

The Tides of Aveiro, Portugal

2011 Icelandic ash cloud

May 22nd 2011 forecast made on May 21st

Forecasting Fire Dynamics

Paradigm shift in Emergency Response

190s could this be forecasted ahead of time?

285s could this be forecasted ahead of time?

316s could this be forecasted ahead of time?

Coupled Fire Mechanisms

Fire dynamics are governed by coupled non-linear processes.

& & & Q (t ) = hc m(t ) = hc m A(t )

Burning rate (per unit area)

from Quintiere, Principles of Fire Behaviour

from Introduction to fire Dynamics, Drysdale, Wiley

Lower order modelling: Two-Zone Model

Upper layer Plume Lower layer

Mass balance lower layer

Flow rate of smoke in plume

Jahn et al., Fire Safety Journal, 2011

Higher order model: Computational Fluid Dynamics

Impossible and HPC

Round-Robin of Fire Modelling

Rein et al., Fire Safety Journal 44, 2009

High Density Instrumentation

20 Heat Flux Gauges

14 Velocity Probes CCTV

Average Compartment Temperature

Abecassis-Empis et al., Experimental Thermal and Fluid Science 32, 2008

Rein et al., Fire Safety Journal 44, 2009

Predictions of smoke layer temperature

Predictions of smoke layer position

Rein et al., Fire Safety Journal 44, 2009

ANALYSIS Forecasts (with positive lead time)

Cowlard et al., Fire Technology, 2011

Inverse modelling is like imitating Sherlock Holmes

Data Assimilation Concept

Invariant values valid

New invariant values

Sudden change of conditions (eg, window breakage)

Source term - Invariants

*= argmin(J()) = the invariant values sought

Jahn et al., Fire Safety Journal, 2011

Tangent Linear Model (TLM)

Leads to a linear system

Jahn et al., Fire Safety Journal, 2011

Zone Model - Typical Fire Scenario

Jahn et al., Fire Safety Journal, 2011

Medium fire (~mattress fire)

Upper layer Plume Lower layer

Layer Height (m)

Medium fire (~mattress fire)

Jahn et al., Fire Safety Journal, 2011

Lead time (s)

Lead time (s)

Lead Time defined as forecast <10% error in upper layer temperature

Lead time (s)