Online Multiscale Conformal Calibration for Response-Defined Wind-Ramp Groups

Wind-ramp events depend on future wind speed, so their group membership is unknown when a forecast is issued. We study MS-CRACP-D, a model-agnostic online calibrator that combines candidate-response inversion, overlapping pooled and site-directional groups, and one-sided partial pooling. Under a finite interval-partition condition, the resulting prediction set is a finite union of intervals. A deterministic controller identity bounds realized group-average miscoverage using a finite-count term and observed projection debt; it does not establish probabilistic conditional coverage at each forecast. A post-review implementation audit identified premature feedback in the archived target-ordered evaluation for overlapping horizons. We therefore reanalyze the frozen-point forecasts using an origin-time event queue and a purged initialization boundary. The corrected R80711 analysis remains a post-review causal reanalysis of previously inspected forecasts. To address the separate questions of turbine-level replication and stochastic point-model sensitivity, we additionally conduct a new end-to-end experiment on ten real turbines: six Kelmarsh turbines and four La Haute Borne turbines. One causal dilated TCN is trained per turbine under five random seeds (50 fits total), with 10-, 60-, and 240-min forecasts evaluated on fixed future test blocks. Across the 30 turbine-horizon streams, MS-CRACP-D marginal coverage ranges from 94.38% to 95.43%, and 29 of 30 streams are at or above 94.5%. On the most severe s3 response-defined ramp groups, MS-CRACP-D attains 84.92–93.27% coverage, versus 9.31–21.36% for GlobalSCP and 3.65–10.73% for StaticRampCP, a gain of 64.56–75.95 percentage points over GlobalSCP. This reliability comes with wider sets, and GlobalSCP has higher marginal coverage on the La Haute Borne aggregate. Across the 30 streams, the mean single-seed MAE and RMSE standard deviations are only 0.00745 and 0.00832 m/s; five-seed averaging improves MAE and RMSE by 3.38% and 2.87% relative to the mean single-seed performance. The resulting evidence supports response-defined ramp protection across two open wind farms, not universal efficiency, geographic transportability to Xinjiang, or joint multi-horizon coverage.