Machine learning-based models to prioritize scenarios in a Quantitative Risk Analysis: An application to an actual atmospheric distillation unit
Résumé
Quantitative risk analysis (QRA) is a systematic methodology to identify, analyze, and calculate risks of an operation or installation of hazardous facilities. One of the first steps of a QRA is the qualitative categorization of the frequency and severity of potential accidents. This task is performed by a group of experts and can be very time and resource-consuming for large-size plants such as oil refineries, which presents several scenarios to look into. This paper presents machine learning (ML) based models to support analysts through the initial stages of a QRA. The proposed approach uses ML classifiers to extract useful knowledge and information from past risk analyses, and thus provide qualitative estimates of consequence and frequency of the accidental scenarios. The approach is demonstrated through a case study concerning an atmospheric distillation unit of an actual oil refinery. The results indicate that the approach is a very promising tool for supporting analysts in the initial stages of QRAs. In addition to reduce time and resources, it can also aid to ensure QRA traceability and reduce variability.