Multi-source uncertainty-aware hierarchical RESCINet residual correction model with conformal adaptive threshold interval for wind turbine condition monitoring

Condition monitoring based on Supervisory Control and Data Acquisition (SCADA) data can track operating states and support wind turbine maintenance decisions. Nevertheless, due to the complex temporal dynamics, variable coupling, local fluctuations, and uncertainty variations, accurate target parameter prediction and reliable threshold interval construction remain challenging. To address these issues, this paper proposes a hierarchical residual-enhanced sample convolution and interaction network (RESCINet) residual correction prediction model and a multi-source uncertainty-driven conformal adaptive threshold interval for wind turbine condition monitoring. In data preprocessing, after Kalman filtering and Pearson-correlation variable screening, random resampling is performed to obtain multiple datasets, thereby mitigating the impact of data uncertainty and enhancing prediction reliability. In the prediction model, a sample convolution and interaction (SCI) mechanism enhanced by a gated residual path is proposed to better retain local sequence information, and candidate features from different network depths are used for hierarchical residual correction of the base prediction. In threshold interval construction, the proposed method constructs a multi-source prediction uncertainty scale from data and model uncertainty, standardizes conformal residuals using this scale, and updates the conformal quantile online to construct the adaptive threshold interval. The proposed method is evaluated on two real-world wind turbine fault cases involving the pitch and converter systems. The results show that the proposed prediction model improves prediction accuracy and that the proposed interval provides earlier alarms than the compared threshold interval methods.