Chemical Process Simulation for Dynamic Risk Analysis: A Steam–Methane Reformer Case Study

Abstract

This paper describes how dynamic simulations of a manufacturing process can be used to construct informed prior distributions for the failure probabilities of alarm and safety interlock systems. Bayesian analysis is used starting with prior distributions and enhancing them with likelihood distributions constructed from real-time alarm data to form posterior distributions, which are used to estimate failure probabilities. The use of alarm data to build likelihood distributions has previously been investigated (Pariyani, A.; et al. AIChE J. 2012, 58, 826−841). Rare-event historical data are typically sparse and have high-variance likelihood distributions. When high-variance likelihood distributions are combined with typical high-variance prior distributions, the resulting posterior distributions naturally have high variances, preventing reliable failure predictions. In contrast with prior distributions obtained by maximizing entropy (Mohseni, T.; et al. AIChE J. 2014, 60, 1013−1026) and those that are based on expert knowledge (Meel, A.; Seider, W. D. Chem. Eng. Sci. 2006, 61, 7036−7056), this paper introduces a new repeated-simulation method to construct informed prior distributions having smaller variances, which in turn leads to posterior distributions with lower variances and more reliable predictions of the failure probabilities of alarm and safety interlock systems. The application of the proposed method is demonstrated for offline dynamic risk analysis of a steam–methane reformer process.