A dynamic risk-informed framework for emergency human error prevention in high-risk industries: A Nuclear Power Plant case study

Abstract

Human reliability analysis (HRA) plays a pivotal role in safety-critical systems, with its methodological evolution currently advancing into the third generation, characterized by dynamic modeling and deeper cognitive processing frameworks. In this study, we propose a novel paradigm extension to HRA, introduced within an emergent operational environment. Specifically, we develop a dynamic risk-informed framework (DRIF) that integrates Bayesian networks (BNs), long short-term memory (LSTM) neural networks, and domain-specific emergency operating procedures (EOPs) to enable real-time evaluation of human error risks during emergency scenarios. The framework employs Bayesian networks to probabilistically model causal relationships among human factors, while LSTM networks dynamically process temporal operational data streams for fault diagnosis. This hybrid architecture synergizes HRA principles with real-time risk propagation mechanisms, thereby enhancing situational awareness and decision granularity under time-critical conditions. To empirically validate DRIF’s efficacy, we implemented it in anomaly mission scenarios for a high-temperature gas-cooled reactor (HTGR). The case study demonstrates the framework’s capability to (1) quantify human error probabilities (HEPs) through probabilistic inference, (2) identify latent risk pathways via backward propagation analysis, and (3) provide prescriptive guidance aligned with EOPs for risk mitigation. The results show that the more precisely later emergency action measures are implemented, the better the accident prevention and control effect during emergencies. This advancement establishes a methodological foundation for next-generation HRA systems in complex engineered systems.