Beyond a Single Risk Score: Posterior Rank Uncertainty in Wildfire Exposure of Transmission Corridors

Yi Wenxing 

Transmission-line wildfire risk management frequently demands segment-level priority triage, whereas records regarding equipment status, power outages and routine inspections remain non-public. This study establishes a reproducible framework relying solely on publicly accessible datasets to conduct retrospective external wildfire exposure ranking for transmission line segments across California. In this workflow, line segments are defined exclusively as wildfire-receiving assets, rather than potential ignition sources, structural failure nodes, outage triggers or liability attribution units.
We compiled a segment-year panel dataset spanning 2017 to 2023 covering the Northern Sierra and Southern Cascades regions, integrating transmission line datasets from the California Energy Commission, CAL FIRE FRAP wildfire perimeters and fire cause datasets, gridMET gridded meteorological summaries, LANDFIRE LF2024 fuel and canopy vegetation datasets, alongside USGS 3DEP topographic data. Fire records labelled as CAL FIRE CAUSE 11 (Electrical Power) are excluded from response variables, effectively separating external wildfire exposure risks from fires directly triggered by electrical facilities.
Under a strict 2022–2023 temporal holdout validation scheme, we benchmarked the Bayesian logistic exposure model against deterministic physical-based scoring methods and conventional machine learning models. Bayesian Laplace and L2 regularized logistic discrimination deliver comparable predictive performance. Accordingly, the core advantages of the Bayesian framework are interpreted via posterior exceedance probability and posterior top-decile ranking probability tailored for rare wildfire event prioritization, rather than general classification accuracy superiority. Given that calibrated probability scores only achieve marginal improvement over simple training-prevalence baselines, posterior rank probability is ultimately recommended as the core decision-support output.