Using a Dynamic Quantitative Risk Assessment System to Manage Day-to-Day Operation Risks

Quantitative Risk Assessment (QRA) is a common safety study used by regulators to understand risks associated with chemical process facilities. Its results are used to provide advice to facility owners with regards to safety as well as to make country-level decisions for land use planning. Common outputs from a QRA include individual risk (IR) per annum and potential loss of life (PLL), both are annualised metrics suitable for regulatory use when assessing safety of facilities that are built to last for several decades. The quantification allows regulators to manage risks across different facility types in an impartial manner, establishing risk level restrictions that all facility owners will have to follow. For example, in Singapore’s context, IR values equal to or greater than 5x10^-5/yr have to be confined within defined facility boundaries, while IR values equal to or greater than 5x10^-6/yr are not allowed to extend into residential areas [1]. If risk values exceed established limits, the regulator could impose on facility owners to deploy active and/or passive protection systems to mitigate critical hazardous event outcomes.

Today, results obtained from a QRA are hardly used by facility owners for purposes other than to meet regulatory requirements, even though it is equally important to manage on-site personnel risks compared to societal risks. The annualised static risk profile does not provide sufficient information that can be used in day-to-day operations. As such, we propose a dynamic QRA system created from modification of the traditional QRA methodology, in order to extract greater value/information for facility operators. This involves the use of data analytics methodologies such as sampling algorithms, surrogate modelling algorithms and error measures, to generate predictive (or surrogate) models that can be coupled with on-site sensors to calculate changing risk profiles for the facility. Variables that will impact on hazardous event outcomes such as wind speed, wind direction, process temperature, process pressure and composition, can be measured with on-site sensors and fed into trained surrogate models to generate “real-time” risk profiles, which can be used by facility operators to make better operational decisions. For example, inspection and maintenance routes can be altered to avoid sending on-site personnel into high risk hotspots, and hot-works (an ignition source) can be temporary suspended when risk in the designated area has risen significantly.

We demonstrate the usefulness of the dynamic QRA system with a case study on a liquefied natural gas (LNG) satellite plant model [3], where consequence analysis has been performed with the use of computational fluid dynamics (CFD) software: Fire Dynamic Simulator (FDS). CFD-based consequence modelling takes into account of geometrical obstructions and hence produces more accurate results than empirical consequence modelling [2], but typically takes hours or days to complete the analysis depending on model complexity. This is where the greatest benefit of using surrogate models can be derived since these models emulating CFD-based consequence modelling can predict hazardous event outcomes in mere seconds.

The example 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 measured at ground level). The surrogate models are built using two distinctively different algorithms: linear nearest neighbours (LNN) and least squares support vector machines (LSSVM) [4], in order to assess their potential for the present application. LNN is a piecewise linear interpolation algorithm, while LSSVM is a global non-linear interpolation model. The training data set is determined with a space filling sampling algorithm, Latin Hypercube Sampling (LHS). Each surrogate model is trained with 100 or 200 data sets and verified against 97 uniquely different data sets. The effect of increasing the number of training data will be evaluated. The accuracy of each surrogate model is determined with multiple error measures including absolute errors, relative errors and root mean square errors (RMSE), and the overall accuracy of using either surrogate model will be presented and discussed.

The focus of the presentation will be on implications of a dynamic QRA system to the industry. How should we convert IR per annum to instantaneous IR for the purposes of the dynamic QRA system? How does this impact the use of common QRA input data? How does exposure time in relation to probits affect the quality of surrogate models being trained? This presentation will serve to highlight, for the first time, the potential of using a dynamic QRA system to enhance safety in day-to-day operations, as well as challenges faced in the creation of the system for large-scale implementation. References

[1] N.E.A. (2017). QRA Criteria Guidelines. National Environmental Agency, Singapore. Retrieved on 01 Apr 2017 from http://www.nea.gov.sg/docs/default-source/anti-pollution-radiation-protection/central-building-planning/revised-qra-guidelines-criteria.pdf

[2] 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.

[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.