2014 CISAP6 Atmospheric Turbulent Dispersion Modeling Methods using Machine Learning Tools

Institut des Sciences des

Risques

Laureta P., Heymesa F., Aprina L., Johanneta A., Dusserrea G., Lapébieb E., Osmontb A.

Atmospheric Turbulent Dispersion Modeling Methods using Machine Learning Tools

6th International Conference on Safety& Environment in Process & Power Industry

Tuesday, April 15, 2014, Bologna, Italy

aLaboratoire de Génie de l’Environnement Industriel (LGEI), Ecole des Mines d’Alès, Alès, France

bCEA, DAM, GRAMAT, F-46500 Gramat, France

Institut des Sciences des Risques (France)Institut des

Sciences des Risques

Modeling Experimental15/04/2014 CISAP6 13-16 April, 2014, Bologna, ItalyInstitut Mines-Telecom2

Context of the study Artificial Neural Networks Methodology Results Improvements & Conclusion

Contents

Atmospheric Turbulent Dispersion Modeling Methodsusing Machine Learning ToolsInstitut des

Sciences des Risques

15/04/2014 CISAP6 13-16 April, 2014, Bologna, ItalyInstitut Mines-Telecom3

ConclusionMethodology


Risques Context Artificial Neural Networks Results


Industrial Site – Flammable/Toxic material storage - Dispersion

Leakage accident

Impact distance < 1 000 m

Exposure time < 1 h

Petrochemical site, Martigues, France


Turbulent Diffusion coefficient estimation






Main goals of this work

1. From quickness to accuracy

Accuracy

Qui

ckne

ss

Gaussian

Integrals

CFD

RANSLES

DNS

Quickness

Accuracy

Closure equations

2. Turbulence modeling

Turbulent diffusion

coefficient calculation

Direct resolution







1. From quickness to accuracy

Gaussian

Integrals

CFD

RANSLES

DNS

Quickness

Accuracy

Turbulent Diffusion coefficient estimation

Developed model

2. Turbulence modeling

Closure equations

Turbulent diffusion

coefficient calculation

Direct resolution

Accuracy

Qui

ckne

ss Turbulent diffusion coefficient forecasting

by Artificial Neural Networks







3. Goals

Developed model

Quickness

Accuracy

Consider cylinder

obstacles

Real experiments

designed

No expert knowledge

required

Near field

Developed model

4. Re>2 x 104

2. Re = 261. Re = 0,16

3. 48<Re<180






Flow around cylinder

1,2,3 from Taneda – 4,5 from Mines Alès

Atmospheric flow: Re > 106

Turbulence modeling is required Unsteady behavior at Re >

2.104

Generally considered as steady in modeling due to random initialization of vortex

Modeling dispersion around cylinder Once wind flow and turbulence

are solved Eulerian: Advection Diffusion

Equation Lagrangian: Particle tracking

5. Shape of flow Behind a cylinder






Artificial Neural Networks (ANN) – Nonlinear phenomenon approximation

Non-linear statistical data modelling tools

Phenomenon database

Inputs Target Output

Neural Network Computed Output Error calculation

Error minimization algorithm

Training Phase






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

Field Experiments

Wind Tunnel Experiments

CFD


Phenomenon database






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

Field Experiments

Wind Tunnel Experiments

CFD


Phenomenon database






Determination of important parameters (Cao, 2007) Position of a plume forecast of continuous standard deviation for gaussian plume

Filter for a gaussian model (Pelliccioni, 2006) Concentrations levels predicted by gaussian model as an input of ANN Other inputs used to refine results are atmospheric conditions parameters Gaussian model improvement

ANN in Atmospheric Dispersion

Conclusions

Three different variables are used: Spatial inputs Atmospheric conditions inputs Case configuration inputs

Database of the phenomenon required






ui and Dt are required Then, ADE can be solve with existing numerical scheme

Using the 2D-Advection Diffusion Equation (ADE) to solve atmospheric dispersion around cylinder

Methodology

ui and Dt forecast using ANN Solving ADE: Finite differences scheme Database characteristics:

Created from CFD model : RANS k- standard with neutral conditions of stability 72 simulations : Diameter m, velocity m.s-1

Domain dimensions: 34 diameters long, 7 diameters large Mesh: from 112 000 to

448 000 nodes Time consuming Sampling is required

to train the ANN

𝜕𝑐𝜕𝑡

+𝑢𝑖𝜕𝑐𝜕𝑥 𝑗

=𝜕𝜕 𝑥 𝑗

(𝐷 𝑡 .𝜕𝑐𝜕 𝑥 𝑗

)+𝑆𝑖+𝑅𝑖

Wind velocity in i directiont TimesC ConcentrationSi Emission source

Ri Reaction

Dt Turbulent diffusion coefficient






Ux, Uy and Dt are each onethe output of an ANN

Inputs variables: Location: polar coordinates Configuration: Diameter Flow conditions: Inlet velocity

Inputs and outputs variables for the ANN

Several ANN models are trained with variations on:

Sampling Number of neurons in hidden layer Parameters initialization

Best model is selected using mean squared error quality indicator.

Training of the ANNDt






Ux, Uy and Dt are each onethe output of an ANN

Inputs variables: Location: polar coordinates Configuration: Diameter Flow conditions: Inlet velocity

Inputs and outputs variables for the ANN

Several ANN models are trained with variations on:

Sampling Number of neurons in hidden layer Parameters initialization

Best model is selected using mean squared error quality indicator.

Training of the ANN






Unlearned test case: D = 12 m ; Uini = 2,5 m.s-1

Coefficient of determination (R²) and FACtor of two (FAC2) are used to qualify the model

Using the ANN for Ux/Uy/Dt determination

Ux Uy Dt

R²: 0,97 FAC2: 0,99 R²: 0,99 FAC2: 0,52 R²: 0,98 FAC2: 0,99

CFD

ANN

CFD

ANN

m.s-1 m.s-1 m2.s-1

CFD

ANN







Flow visualization

CFD

ANN

Velocity vectors

CFD

ANN

Streamlines






Wind flow and Turbulent diffusion coefficient are used to solve the ADE Finite differences are used Explicit resolution for advection and diffusion terms Stability criteria has to be set :

Courant number is used for the advection terms:

Diffusion terms has to respect:

Minimum is selected

Cylinder obstacle is detected and convert on a rectangular mesh Boundary conditions are set as in CFD model Comparison is made from same initial concentrations

CFD Wind flow and Dt are interpolated on the new mesh

ANN Wind flow and Dt are calculated on the center ofthe mesh cells

Using the ADE








CFD

ANN








CFD

ANN







CFD

ANN


Computing time: Flow field and Dt by ANN model: less than 2 seconds Flow field and Dt by CFD turbulence model: from 20 minutes to 1 hour But with different resolutions

Advection diffusion equation ~3 minutes for 1 minute in simulation time With spatial resolution of 0.5 m Optimization has to be made

Computer used : Classical workstation Processor: Intel® Core™2 Duo CPU: E7500-2,93 GHz RAM: 4 Go Windows 7 Professionnal CFD software: Ansys® Fluent 14 Academic Research

Wind flow and turbulent diffusion coefficient modeling is very fast Accuracy is evaluated through CFD comparison Model has to be confront to experimental data Turbulent dispersion is correctly modeled around a cylinder Data needed are only diameter and inlet velocity to compute

turbulence in neutral stability conditions ANN in combination with ADE resolution act as a grey box.






Conclusion

Quickness

Accuracy

Consider cylinder

obstacles

Real experiments

designed

No expert knowledge

required

Near field

Developed model

Experimental data acquisition are needed: Comparison with current model Training on real life data

Future work will be focused on dispersion over multiple obstacles Tridimensional modeling of the flow field and Dt will be implement Numerical optimization has to be done

Perspectives

This research was supported by the CEA: French Alternative Energies and Atomic Energy Commission

Acknowledgements


Risques

Laureta P., Heymesa F., Aprina L., Johanneta A., Dusserrea G., Lapébieb E., Osmontb A.


6th International Conference on Safety& Environment in Process & Power Industry

Tuesday, April 15, 2014, Bologna, Italy

aLaboratoire de Génie de l’Environnement Industriel (LGEI), Ecole des Mines d’Alès, Alès, France

bCEA, DAM, GRAMAT, F-46500 Gramat, France