Certified Group-Robustness Bounds for Artificial Intelligence Climate Downscaling Under Transboundary Station Outages: Coverage Guarantees and Hydropower Dispatch Reliability

Artificial intelligence downscaling models for climate variables are typically validated on complete station records, yet transboundary observing networks suffer outages that remove entire groups of stations simultaneously, so that the residual training distribution is neither exchangeable nor representative of the deployment domain. The framework partitions the calibration set by station group, constructs a worst-group coverage functional, and derives a tight upper bound on the coverage gap between the pooled predictor and any group-conditional predictor under a bounded group-shift assumption. The bound is shown to be sharp and to reduce to the standard conformal guarantee when group exchangeability holds. The framework is applied to NASA POWER monthly reanalysis air temperature at 2 m for Lagos, Nigeria (6.52° N, 3.38° E), 2000–2023, comprising 312 monthly observations with a mean of 26.414 °C and a warming trend of 0.1553 °C per decade. The certified intervals are propagated into a hydropower dispatch reliability calculation, and the analysis shows how group-robustness certification translates into a quantifiable reduction in the risk of under-forecasting reservoir inflow temperature constraints. The contribution is a coverage certificate that remains valid under adversarial group removal, together with an explicit statement of the conditions under which that certificate is conservative.