Extreme weather events and operational failures now cause longer, more widespread, and far less predictable power outages. Power system resilience planning therefore requires probabilistic models that can quantify outage duration and customer impact across the full spectrum of disruption scenarios, weather-driven and operational alike, within a unified, spatially disaggregated framework. Existing data-driven approaches treat these two event classes separately, depend on fixed-effect regression that cannot accommodate regional data sparsity, and provide limited uncertainty quantification for infrastructure decision support. This paper develops a Bayesian hierarchical model that jointly characterizes weather-related and non-weather-related distribution-level power outages using a decade of EAGLE-I outage records (2014–2023) aligned with National Weather Service meteorological metadata (1986–2023). The model captures state-to-county geographic heterogeneity through partial pooling and encodes season–event interaction patterns. Both features combine to produce probabilistic estimates even for data-sparse counties and rare event categories. Joint modeling of outage duration and customer impact within a single probabilistic structure yields full posterior uncertainty quantification that directly supports resilience benchmarking across geographic regions, scenario-based capital planning, imputation of incomplete outage records, and severity-aware resource allocation for utilities and grid operators. Posterior diagnostics confirm stable convergence ( R ̂ ≤ 1 . 01 ) across all parameters; the hierarchical structure yields interpretable estimates at state, county, and seasonal resolution that reveal region-specific and event-specific drivers of outage behavior. The results demonstrate that weather-driven disruptions require qualitatively different response protocols from operational outages, that event severity scales non-uniformly with customer impact across event classes, and that seasonal hardening investments can be guided by posterior interaction effects rather than blanket infrastructure upgrades. • A unified Bayesian model jointly characterizes weather and non-weather grid outages. • State-to-county partial pooling yields calibrated estimates for data-sparse counties. • A novel NWS–EAGLE-I composite scoring pipeline produces empirically informed priors. • Posterior uncertainty quantification supports benchmarking, planning, and imputation.
Chaudhary et al. (Wed,) studied this question.