Making Sense of Dynamic PSA Results: A Hybrid Approach

A dynamic PSA in form of an integrated deterministic probabilistic safety analysis (IDPSA) combines the enhanced realism of a deterministic safety analysis (DSA) with the advantages of a probabilistic safety analysis (PSA). The GRS software tool MCDET (Monte Carlo Dynamic Event Tree) for dynamic PSA allows to analyse and quantify the influence of aleatory and epistemic uncertainties on the behaviour of dynamic systems over time. It can be used both to identify unforeseen accident sequences as well as to quantify the dependencies between different end state scenarios and the respective uncertain input parameter(s).
The effects of the high-dimensional parameter space induced by variations of the system state and the timing of events are simulated and represented using a Monte Carlo approach in combination with the dynamic event tree simulation. This results in large samples of event trees and time-dependent scenarios requiring state-of-the-art meth-ods of data analysis for analysing the huge amount of data generated.
This paper introduces how data analysis and machine learning can be used together with domain knowledge to extract and condense the relevant information, to estimate safety margins and to determine the most significant discrete and continuous parameters. Furthermore, it is outlined how various techniques can be combined in an ex-ample application of an accident during mid-loop operation in the ADAMO project (Application of Advanced Dy-namic PSA Methods for Assessing the Effectiveness of Human Actions for Accidents in Mid-Loop Operation). It also illustrates how interactive data visualization can be used to understand system processes and component inter-actions leading to the time series of dependent variables derived in an IDPSA.

Keywords: Probabilistic Safety Analysis (PSA), IDPSA, Dynamic PSA, Machine learning, Mid-loop operation.