Resilience-Based Process Upset Event Prediction Analysis for Uncertainty Management Using Bayesian Deep Learning: Application to a Polyvinyl Chloride Process System

Abstract

There are uncertainties involved in the risk assessment of process systems operations. Also, systems are complex and deteriorate gradually with time or due to exposure to expected or unexpected disturbances and events. Questions such as “what is the frequency of a process upset?” and “can we predict incidents?” have yet to be explored and answered. With the use of the Process Resilience Analysis Framework, this work presents a resilience-based approach to managing uncertainties to better predict process upsets. Prior specification on uncertain parameters is assumed based on historical data. Popular sampling and Bayesian techniques such as Markov chain Monte Carlo simulation and mixture modeling are used for posterior inference on the parameters. The application of the predictability assessment for uncertainty management is demonstrated using a polyvinyl chloride process system. A total of three types of uncertainties (cooling medium temperature, agitator failure, and reactants charging) are considered. It is concluded that with the use of resilience metrics data, the variance of statistical parameters can be updated, leading to high-probability regions of the parameter space responsible for the observed data. This helps the risk assessors make more-accurate and informed process risk decisions.