A new approach for deep prediction of urban complex system risk process during natural disasters

In recent years, advancements in weather forecasting systems have led to increased accuracy. Despite more accurate disaster input conditions, predicting the risk evolution process of the urban complex system remains an unresolved issue which represents a vulnerable link. Currently, there are numerous methods for risk prediction. However, a universally applicable approach and fundamental model that can dynamically predict the urban risk process under varying disaster input conditions have not been established yet. To address these challenges, we propose an event graph model within the framework of the extended risk concept. Furthermore we introduce the theory of Directed Markov Random Field to construct an Urban Spatio-temporal Risk Process model (USTRP), which enables the dynamic forecasting of risk process. The USTRP model can address basic problems in application such as identifying the most or more probable event chains, calculating the node marginal distribution, and determining the first hitting time under different disaster conditions. Moreover, to improve computational efficiency, we leverage the characteristics of the USTRP and present a sparse low-entropy approximate direct inference algorithm (SLEADIA) while proving its convergence. Finally, we apply this model to a hypothetical case. We analyze the medical service acquisition capabilities of nursing homes and the leakage risks of chemical storage tanks under varying flood conditions, demonstrating the computational efficiency advantage of the proposed SLEADIA.