Uncertainty-Calibrated Residual Conformal Monitoring of Wind Turbine SCADA Data for Cross-Asset Anomaly Detection Under Distribution Shift

Abstract

Wind turbine SCADA anomaly detectors are commonly calibrated using data from the same asset or development period, whereas deployment requires transfer across turbines with different operating distributions. This mismatch can produce optimistic thresholds, excessive false alarms, and misleading generalization estimates. This study proposes UC-RCF, an uncertainty-calibrated residual conformal framework integrating nonlinear multi-output normal behavior modeling, embargoed blocked cross-fitting, uncertainty-normalized residuals, operating support assessment, channel-wise conformal evidence, reflected cumulative criticality, and cross-asset alarm calibration. Evaluation followed a nested leave-one-turbine-out, asset-disjoint protocol on CARE v6, comprising 95 monitoring cases from 36 turbines across three wind farms, including 45 anomalous and 50 normal cases. UC-RCF achieved a pooled outer-fold CARE score of 0.5903, fault coverage of 0.3920, weighted earliness of 0.2423, event-level reliability of 0.5760, and normal-case accuracy of 0.8706. It detected 25 anomalous cases and generated alarms in 18 normal cases. A farm-stratified paired turbine-cluster bootstrap estimated a CARE improvement of 0.0349 over the initial full configuration, with a 95% percentile interval of 0.0055–0.0671 and a bootstrap probability of improvement of 0.991. These estimates remain exploratory because the effective resampling units comprise only 36 physical turbines from three wind farms. Mahalanobis monitoring achieved a higher CARE score of 0.6012, detecting 21 anomalous cases while generating alarms in nine normal cases. UC-RCF therefore provided broader fault coverage and four additional anomalous-case detections but incurred a higher false-alarm burden, demonstrating a sensitivity–reliability trade-off rather than universal detector dominance. Ablation analysis identified farm-normalized residual magnitude as the strongest case-level discriminator, with a receiver operating characteristic area under the curve of 0.854. Heteroscedastic uncertainty scaling and operating support adjustment did not consistently improve CARE or raw discrimination; the auxiliary framework components instead provide mechanisms for uncertainty characterization, score comparability, temporal persistence, and calibration auditing. At the CARE-optimal nominal case-level false-alarm budget of α=0.30, the empirical normal-case false-alarm rate was 0.36. Because temporal dependence and cross-asset distribution shift can violate exchangeability, α is interpreted as an operational calibration target rather than a theoretically guaranteed case-level error bound. Overall, UC-RCF provides an interpretable and auditable framework for investigating residual evidence, operating support shift, temporal persistence, and calibration reliability under cross-asset deployment.