Integrating dynamic Bayesian network and physics-based modeling for risk analysis of a time-dependent power distribution system during hurricanes

Hurricane is one of the major natural hazards that bring significant damages and failures to the power distribution system for many coastal regions. For better decision-making, pre-hazard maintenance and post-hazard restoration of the power distribution system can be improved through the risk analysis of the distribution system. However, as a natural material, wood deteriorates with time. Therefore, the performance of the power distribution system could vary significantly during their life cycles. To quantify the time-dependent nature of the existing power distribution system due to material deterioration, a dynamic Bayesian network (DBN) is implemented for the risk analysis of a power distribution system against hurricanes. To capture the failure probabilities of the poles in the power distribution system, fragility surfaces for the poles are generated through physics-based fragility analysis, incorporating physical properties of the pole-wire components and the pole-wire system topology. The failure rate of the distribution system is obtained through the Bayesian network (BN) at each time step. With the observed failure data of poles, the material properties and the fragility surfaces of the components in the power distribution system are updated through Bayesian inferring. More specifically, the particle filter (PF) algorithm is utilized. To efficiently save computational cost in BN reasoning, the Gaussian process (GP) algorithm is adopted to build surrogate models to predict poles’ behaviors. Predictors are selected from the feature space to further reduce computational time and to enhance the prediction performance of the surrogate models through the automatic relevance determination (ARD) method. The failure rate is underestimated by 11% when the system deterioration process is ignored given a ten-year interval of hurricane events.