(56e) Large Scale Surrogate Modelling for Enhancement of Consequence Modelling of Industrial Fires

Loy Yoke Yuan^a,b,
Gade Pandu Rangaiah^a and Lakshminarayanan Samavedham^a

^aDepartment of Chemical and
Biomolecular Engineering, National University of Singapore, Singapore 117585

^bLloyd’s Register Global Technology
Centre Pte Ltd, 1 Fusionopolis Place, #09-11 Galaxis, Singapore 138522

  Abstract

Quantitative
Risk Assessment (QRA) traditionally employs empirical consequence models to
calculate damage values resulting from a hazardous event, such as the radiation
flux of industrial fires. With rapid
advancements in computational technology, it is now possible to perform
consequence modelling with computational fluid dynamics (CFD). CFD-based
consequence modelling is arguably more accurate than traditional methodologies
since it takes into account the effects of geometrical obstructions [1]. It
also avoids some assumptions inherent with the use of empirical models for
consequence modelling. However, the use of CFD is very costly due to the long
simulation time, particularly when multiple events have to be modelled for the
purpose of a QRA [2]. As a result, this becomes a barrier for CFD-based QRA to replace
traditional QRA methods in the market. In this paper, it is proposed that the
value of CFD-based consequence modelling for
industrial fires can be increased with the use of surrogate models so as
to enhance information gain and possibly reduce the number of simulations
required to perform a full CFD-based QRA. This is a novel application of data
analysis methodologies for QRA.

The
present work investigates the accuracy of surrogate models generated using
local linear interpolation (LLI) and support vector regression (SVR) algorithms,
for industrial fires. A liquefied natural gas
(LNG) satellite plant model [3] is used as the case study, with Fire Dynamics
Simulator as the CFD software and SUMO toolbox [4] for the surrogate modelling.
The study involves two input variables (wind speed and heat release rate per
unit area of pool fire) and one output variable (time-averaged net radiation
flux collected at ground level). The output variable is collected as a maximum
time-average value throughout the simulation time, time-average value at 50
seconds and time-averaged value at 100
seconds.  Each surrogate model is trained by a fixed 100 input points designed
with Latin Hypercube Sampling (LHS), and verified with 97 verification points
selected differently from the LHS design. The accuracy of each surrogate model is
measured with root mean square error (RMSE), and the overall accuracy of each
algorithm (LLI and SVR) is presented. Our study will serve to highlight the
potential of surrogate modelling for realistic consequence modelling of industrial fires as well as challenges faced in
surrogate modelling of complex systems.

  References

[1]
Hansen, O.R., Davis, S.G. and Gavelli, F. (2012). Benefits of CFD for onshore
facility explosion studies. 8^th Global Congress on Process Safety,
Houston Texas, 1-4 April 2012.

[2]
Kajero, O.T., Thorpe, R.B., Chen, T., Wang, B. and Yao, Y. (2016). Kriging
meta-model assisted calibration of computational fluid dynamics models. AIChE
Journal. DOI 10.1002/aic.15352.

[3]
Sun, B., Guo, K. and Pareek, V.K. (2014). Computational fluid dynamics
simulation of LNG pool fire radiation for hazard analysis. Journal of Loss
Prevention in the Process Industries (29), 92-102.

[4]
Gorissen, D., Couckuyt, I., Demeester, P., Dhaene, T. and Crombecq, K. (2010).
A surrogate modeling and adaptive sampling toolbox for computer based design.
Journal of Machine Learning Research (11), 2051-2055.