loss prevention 2013 - atmospheric dispersion modelling by cellular automata and artificial neural...

Loss Prevention 2013 – 12-15 Mai - Firenze

Near Field Atmospheric Dispersion Modelling on an Industrial Site Using Combination of Cellular Automata and Artificial Neural Networks

-

Pierre Lauret, Ph.D. Student, 2nd year

Frédéric Heymesa, Laurent Aprina, Anne Johannetb, Gilles Dusserrea, Laurent Munierc, Emmanuel Lapébiec

a: Ales Shool of Mines, Franceb: Commissariat à l’Energie Atomique et aux Energies Alternatives

Institut des Sciences des Risques

Context of the study Machine learning tools Methodology Results Improvements Conclusion

Summary

04/12/2023 2Loss Prevention 2013 – 12-15 Mai - Firenze

Institut des Sciences des

Risques

Near Field Atmospheric Dispersion Modelling –Cellular Automata and Artificial Neural Networks

Context of the study Industrial site Existing models

Machine learning tools Methodology Results Improvements Conclusion

Summary



Risques


Context of the study Machine learning tools

Cellular Automata - CA Artificial Neural Networks - ANN Combination

Methodology Results Improvements Conclusion

Summary



Risques


Context of the study Machine learning tools Methodology

CFD case Database creation Scalar transport equation

discretisation ANN - Learning ANN – Optimisation CA-ANN – using the model

Results Improvements Conclusion

Summary



Risques


Context of the study Machine learning tools Methodology Results

Test cases Discussion

Improvements Conclusion

Summary



Risques


Context of the study Machine learning tools Methodology Results Improvements Conclusion

Summary



Risques


Machine learning MethodologyContext

Impact distance < 1 000 m

Exposure time < 1 h

Industrial site, Martigues, France

Leakage accident

Industrial Site – Flammable/Toxic material storage - Dispersion

Results Improvements Conclusion

Near Field Atmospheric Dispersion Modelling - Cellular Automata and Artificial Neural Networks


Institut des

Sciences des

Risques

Methodology development Experimental validation Uncertainty Quantification and Limits of used Aim : Providing operational model with several

goals :

Gaussian models (Solving the Advection-Diffusion Equation)

Integral models Computational Fluid Dynamics (CFD) models

Reynolds Averaged Navier-Stockes equations (RANS) Large Eddy Simulation (LES) Direct Numerical Simulation (DNS)

Quickness

Complete ness

Considera tion of

obstacles

Real experiments

designed

Short time dispersion

Near field

Developed modelQuickness

Completeness

Existing models / Developped model

From quickness to completeness CA-ANN


Machine learning MethodologyContext Results Improvements Conclusion

Near Field Atmospheric Dispersion Modelling - Cellular Automata and Artificial Neural NetworksInstitut

des Sciences

des Risques

Cellular Automata (CA) – Spatial and temporal representation

tt+dt

Transition rules example

tt+dt

« 0 » States

« 1 » States

Monodimensional Cellular Automata Example

t0

t1

t2

t3

t4

t5

Cellular Automata evolution

Cellular Automata are designed by : Regular mesh Finite possible states for each cell Transition rules

State evolution is done by applying local and deterministic rules, uniformly and synchronously for all cells.




des Sciences

des Risques


tt+dt


tt+dt

« 0 » States

« 1 » States


t0

t1

t2

t3

t4

t5







des Sciences

des Risques

Artificial Neural Networks (ANN) – Non linear phenomenon approximation

Non-linear statistical data modelling tools

Phenomenon database

Inputs Target Output

Neural Network Computed Output Error calculation

Error minimization algorithm




des Sciences

des Risques

Non-linear statistical data modelling tools Parameters modification to minimize the ANN error Database of the phenomenon required

Phenomenon database

In Field Experiments

Wind Tunnel Experiments

CFD




des Sciences

des Risques

Artificial Neural Networks (ANN) – Non linear phenomenon approximation

Cellular Automata Artificial Neural Networks ruled (CA-ANN)

Cellular Automata

Spatial and temporal representation

Non linear phenomenon approximation

Navier-Stokes equations emulation

Cellular Automata

Artificial Neural Networks

Cellular Automata Artificial Neural Networks ruled

Artificial Neural Networks




des Sciences

des Risques

Free Field Methane Atmospheric dispersion Puff evolution

Sequence: ejection of CH4, self evolution until exiting the numerical domain Each time step, case is saved (Velocity field/ CH4 Concentration) Time step duration is related to Courant number

Intervals: Wind velocity: 2-20 m.s-1

Ejection velocity: 2-20 m.s-1

CH4 Mass Fraction: 0.2-1 Time step : 20

Methane ejection with initial velocity – RANS k-ε realizable model




des Sciences

des Risques

CFD Database constitution

Wind:velocity inlet

Wind:velocity inlet

CH4: velocity inlet

Formatting data

Starting from advection-diffusion equation (Navier-Stokes) :




des Sciences

des Risques

Ux

Uy

t0+dt

C

t0

For each cell i,j:

Formatting data





des Sciences

des Risques

Ux

Uy

t0+dt

C

t0

For each cell i,j:

Formatting data





des Sciences

des Risques

For each cell i,j:

Formatting data





des Sciences

des Risques

TargetANN Inputs

Collected Computed

Training and optimization

ANN Parameters Initialization

Sampling Database

ANN Architecture

C0 initialization from an existing case




des Sciences

des Risques

Simulating a known case and evaluating the CA-ANN

Iterative algorithm ofthe Cellular Automata

Artificial Neural Network ruled (CA-ANN)

t0 : Initialization





des Sciences

des Risques

Using and evaluating the CA-ANN



t0 : Initialization





des Sciences

des Risques




t0 : Initialization





des Sciences

des Risques




t0 + dt…

Unlearned case simulation and evaluation




des Sciences

des Risques

Initialization: using a CFD case (Initial Mass Fraction: 0.89; Wind velocity: 10.2 m.s-1)Mass conservation trough time steps:

Mass evolution for CFD simulation and CA-ANN model

Tota

l mas

s

Time steps

Mass (CA-ANN)

Mass (CFD)





des Sciences

des Risques

Initialization: using a CFD case (Initial Mass Fraction: 0.89; Wind velocity: 10.2 m.s-1)Mass conservation trough time stepsUncertainty visualisation: Model Vs CFD reality

Model concentrations Vs CFD concentrations: at time step

CA-A

NN

con

cent

ratio

ns (k

g.m

-3)

CFD concentrations (kg.m-3)





des Sciences

des Risques

Initialization: using a CFD case (Initial Mass Fraction: 0.89; Wind velocity: 10.2 m.s-1)Mass conservation trough time stepsUncertainty visualisation: Model Vs CFD reality

Model concentrations Vs CFD concentrations: at time step Absolute error at time step

CA-A

NN

con

cent

ratio

ns (k

g.m

-3)

CFD concentrations (kg.m-3)

and CA-ANN concentrations(>0: over estimation ; <0 under estimation)

dx

dy

Improvements

Wind field determination Experimental (wind tunnel) and numerical database

(CFD) Cylindrical obstacles (storages) ANN modelisation Several obstacles dimensions Several inlet velocity

Coupling wind field model (ANN) and atmospheric dispersion model (CA-ANN)

Preliminary tests



des Sciences

des Risques

Conclusions

Existing forecasting models are slow but accurate or fast but not enough appropriate. A method using Cellular Automata and Artificial Neural Networks is presented. Comparison with CFD software gives good agreement and faster processing. Computing optimization is not yet done. CA-ANN lightly overestimates the higher concentrations. A decay in the determination coefficient trough time steps is noted. ANN training optimization process is in progress.




des Sciences

des Risques

Loss Prevention 2013 – 12-15 Mai - Firenze

Institut des Sciences des Risques

Near Field Atmospheric Dispersion Modelling on an Industrial Site Using Combination of Cellular Automata and Artificial Neural Networks

-

Contact: [email protected]

loss prevention 2013 - atmospheric dispersion modelling by cellular automata and artificial neural...

Documents

mai firenzeinstitut

mai firenze2

mai firenze3

mai firenze4

mai firenze5

mai firenze6

mai firenze7

mai firenze8