(560c) Development of Computationally Efficient Dynamic Model to Estimate Consequence of Rare Events

In the chemical process industry, rare events such as toxic release, fire, and explosion incidents are widely studied due to their high environmental, economic, and social impact [1]. To analyze various aspects of high-consequence rare events, computational fluid dynamics (CFD) models have been widely used, for example, to obtain the concentration profile of materials released during rare events. However, CFD models are computationally intensive in nature as they solve a set of coupled differential equations [2]. Therefore, several developments have been made in building computationally efficient models. The existing computationally efficient models in the field of high-consequence rare events are temporally static and thus, they are not able to describe process dynamics [3, 4]. However, it is crucial to compute the temporal concentration profile of released materials at multiple locations for consequence modeling of rare events. Further, the consequences depend on various parameters (e.g., orifice diameter) which may vary with different release scenarios. To capture the effect of parameters on the concentration dynamics, it is not affordable to develop a new model for every parameter value [5]. Hence, for consequence estimation of rare events, there is a need to develop a computationally efficient dynamic model which is robust with respect to parameter change.

To address these challenges, this work develops a dynamic k-nearest neighbor (kNN)-based parametric reduced-order model (PROM), which can replace computationally demanding CFD models for consequence modeling and handle any changes in parameters. In this work, multivariable output-error state space (MOESP) algorithm was selected to construct the dynamic ROM due to its ease of implementation, and a kNN algorithm was employed among various machine learning algorithms because of its good performance in modeling a physical system with a limited availability of data (e.g., rare-event data). Specifically, the proposed approach interpolates local (with respect to parameters) ROMs constructed for a range of parameters in two steps. First, local ROMs are constructed using the MOESP algorithm. Then, the concentration profile for a new parameter value is obtained by interpolating the concentration profiles obtained from k-nearest local ROMs. Next, the obtained concentration profile is used with a well-developed dose-response model to estimate consequences. The effectiveness of the proposed kNN-based PROM was demonstrated through a case study of supercritical carbon dioxide release rare event. To conclude, this work contributes towards the development of consequence models by proposing a computationally efficient dynamic model capable of quantifying the effect of parameters.

Keywords: parametric reduced-order model; k-nearest neighbor model; consequence estimation; rare events

References:
1. Khan, F., Rathnayaka, S., Ahmed, S., 2015. Methods and models in process safety and risk management: Past, present and future. Process Safety and Environmental Protection 98, 116-147
2. Zhang, B., Chen, G.M., 2010. Quantitative risk analysis of toxic gas release caused poisoning: a CFD and dose-response model combined approach. Process Safety and Environmental Protection 88, 253-262
3. Jeon, K., Yang, S., Kang, D., Na, J., Lee, W.B., 2019. Development of surrogate model using CFD and deep neural networks to optimize gas detector layout. Korean Journal of Chemical Engineering 36, 325-332.
4. Sun, Y., Wang, J., Zhu, W., Yuan, S., Hong, Y., Mannan, M.S., Wilhite, B., 2020. Development of consequent models for three categories of fire through artificial neural networks. Industrial & Engineering Chemistry Research 59, 464-474.
5. Benner, P., Gugercin, S., Willcox, K., 2015. A survey of projection-based model reduction methods for parametric dynamical systems. Society for Industrial and Applied Mathematics 57, 483-531.