Allison Reilly, DOGEE
A number of models have been developed to estimate the spatial distribution of infrastructure impact during a natural hazard event. For example, statistical approaches have been developed to estimate the percentage of customers without power due to a hurricane, with the estimates made at a local geography level such as census tracts. While some statistical infrastructure performance models use extensive covariate data that captures a significant amount of the spatial information, others use a limited number of covariates to enhance model simplicity and to reduce the cost and time associated with obtaining this data. However, in these simpler models, the omitted covariates result in loss of spatial information in the model, leading to a situation in which predictions from adjacent regions are more dissimilar than would be expected. In this paper, we develop a tree-based statistical mass-balance multiscale model to smooth the outage predictions at granular levels by allowing spatially similar areas to inform one another with the goals of: (1) reducing spatial error in simplified prediction models and (2) yielding estimates at other levels of aggregation in addition to the native model resolution. We use a generalized density-based clustering algorithm to extract the hierarchical spatial structure. The “noise” regions (i.e., those regions located in sparse areas) are then aggregated using a distance-based clustering approach. We demonstrate this approach using outage predictions from Hurricane Ivan and develop outage prediction maps at different levels of granularity.