Evidence-Gated Deployment of Cross-Farm Wind Power Forecasts Under Negative Transfer and Sensor Uncertainty

Abstract

Cross-farm adaptation can improve wind-power forecasting when target labels are scarce, but the same update can also cause negative transfer. We evaluate DART-Guard v2.1 as an empirically calibrated, reference-conditional deployment policy that separates proposal selection from release authorization and calibrates the release statistic after proposal-based candidate selection. Five public 10-min wind-farm datasets support internal, external-replication, and legacy boundary analyses. Target formulation is frozen from source-side or chronologically prior validation evidence, headline deployment uses non-overlapping farm-episodes, and five model seeds are averaged before each operational decision. Under the prespecified 0.5 pooled-source/0.5 persistence reference, an uncalibrated post-selection Student-t screen produced a 7.56% empirical global-null release rate. Leave-one-episode-out matched-window selection-aware calibration reduced this to 4.88%, close to the nominal 5%. Across 20 farm-episode replays, the recalibrated policy authorized candidate deployment in 5 episodes, releasing 1 of 6 harmful opportunities and retaining 4 of 14 beneficial opportunities. Deployment behavior changed materially with the declared reference, adaptation benefit remained route-, budget-, and loss-dependent, and a source-only tuned lightweight Patch Transformer did not outperform persistence in any of eight route-horizon settings. These results support auditable conditional release with explicit coverage-risk-reference trade-offs rather than a guarantee of safe transfer.