Abstract
<jats:p>Bangladesh faces rising intraplate seismicity beneath a soft-soil capital governed by an outdated 1979 seismic-hazard map. During earthquakes, power distribution networks directly cause casualties through three mechanisms: electrocution from downed conductors, fault-ignited fires, and the disruption of critical life-support loads. Standard resilience models minimize energy-not-supplied (ENS), ignoring these direct human impacts. To address this gap, we propose the first casualty-weighted, risk-averse, and distributionally robust network design. Our two-stage stochastic mixed-integer program minimizes expected severity-weighted casualties—mapped to HAZUS injury levels—using component hardening, distributed-generation siting, and seismic-shutoff switching under a LinDistFlow model. To hedge against the deep uncertainty of outdated hazard data, the model employs Conditional Value-at-Risk (CVaR) alongside total-variation and Wasserstein ambiguity sets featuring decision-dependent uncertainty, solved via column-and-constraint generation. Tested on a modified IEEE 33-bus feeder calibrated to a decade of national demand, the casualty-focused objective reduces expected fatalities from 25.7 to 0.78. An equal-budget ENS design performs poorly, as it cannot perceive residual electrocution and fire risks. Crucially, we demonstrate that rapid de-energization and backup generation significantly outperform traditional structural hardening in minimizing human casualties.</jats:p>