Network-aware multi-step hazard prediction using temporal knowledge graphs: A chemical industry case study

Abstract

Proactive hazard prediction in complex industrial environments like the chemical sector is critical yet challenging due to dynamic, interconnected risks often overlooked by traditional methods. Existing data-driven approaches frequently fall short by failing to model evolving temporal dependencies and multi-step risk propagation across diverse hazard relationships. To overcome these limitations, this study introduces the Temporal Knowledge Graph-Autoregressive Multistep Prediction Model (TKG-AM). Our core innovation lies in representing dynamic, multi-relational hazard data using Temporal Knowledge Graphs (TKGs) and coupling this rich representation with an autoregressive deep learning engine specifically designed for accurate multi-step forecasting, providing crucial lead time for interventions. Validated on extensive hazard records from a chemical industrial park in Ningxia, China, TKG-AM demonstrated strong predictive power, achieving a direct hit rate (Hits@1) of 58.5 % and top-ten accuracy (Hits@10) of 67.3 %. Our analysis revealed the network's small-world properties, facilitating rapid risk diffusion, and identified 75 critical bridging nodes central to information flow. We further analyzed how network topology and specific relationship types impact prediction accuracy, finding, for instance, that inter-community predictions are inherently more challenging. To enhance practical application, we developed a data-driven prediction score threshold enabling risk prioritization (e.g., scores >20 yielding >90 % accuracy). These integrated findings validate TKG-AM as a robust and insightful methodology, offering significant improvements in the efficiency, specificity, and strategic targeting of hazard prevention and differentiated risk management efforts in the chemical industry.