Multi-hazard probabilistic safety assessments using Bayesian networks – A framework and demonstration for integrating technical and human risk

• Introduces a Bayesian network framework for multi-hazard risk integration. • Demonstrates with a realistic case of a nuclear power plant exposed to seismic and flood hazards. • Embeds technical and human reliability models as subnetworks. • Compares unified vs object-oriented integration of subnetworks. • Links accident probability to factors like operator stress levels and flood defence stability. Despite the advantages of using Bayesian networks for probabilistic risk assessment, adoption in practice has been limited due to the lack of realistic, facility-scale studies. Scaling up from systems to facility-level safety assessments poses challenges in (i) integrating external hazards and their cascading effects, and (ii) resolving non-homogeneity of various technical and human reliability models. The novelty of the study is in formalising risk integration using Bayesian networks, at facility scale, and demonstrating its effectiveness in addressing associated challenges. A Bayesian network-based multi-hazard risk framework is introduced and demonstrated for a nuclear power plant subject to flooding and earthquake hazards, capturing dependencies among hazards and consequences. Individual reliability models – conventionally extraneous to facility-wide risk models – are included as subnetworks by using Bayesian network-based surrogate models for technical systems and a Bayesian networks approach for human reliability modelling. Two approaches are used for subnetwork integration – object-oriented and unified Bayesian networks. The unified approach allows for prediction, diagnostics and inter-causal reasoning since Bayesian inference is bi-directional. Conversely, in the object-oriented approach, diagnostics are limited to within individual subnetworks and as a consequence the model can potentially neglect dependencies between objects. However, the object-oriented model requires only 50 % of the computational memory and consumes less than 25% of the runtime as the unified network, while improving visual clarity of the risk model. The model reveals key insights – for example, variations in operator stress or available response time during a hazard event can result in up to a 77 % change in top event probability – demonstrating its effectiveness in capturing critical relationships in complex, facility-scale risk scenarios. These findings can be used to suitably allocate resources towards risk mitigation and plant safety management.