CrossSTLLM: A continually adaptive spatiotemporal large language model for anomaly detection and fault localization in wind turbines with diverse distribution

Wind turbine SCADA data varies widely across sites and over time, resulting in cross-spatiotemporal heterogeneity and evolving operating conditions that can degrade traditional anomaly detectors. To address these challenges, a novel anomaly detection and fault localization framework is proposed, leveraging a continually adaptive cross-spatiotemporal large language model (CrossSTLLM) and a statistic indicator. CrossSTLLM adapts to spatiotemporal heterogeneity and concept drift through continual learning, thereby capturing evolving operational patterns and integrating new site-specific data distributions. During online operation, a condition monitoring indicator based on Hotelling's T2 statistic is proposed for interpretable anomaly detection and fault localization. The multivariate outputs of CrossSTLLM are subjected to Hotelling's T2-based monitoring, yielding a unified deviation score for each time window. When this T2 statistic exceeds a predefined control limit, an anomaly is declared and the contributions of individual sensors to the statistic are analyzed to localize the fault. In this manner, the anomalies are flagged and the SCADA variables driving abnormal behavior are identified. The approach is validated on SCADA data from two real wind farms, where CrossSTLLM provided early fault alerts 27 and 26 days before maintenance events. Quantitatively, CrossSTLLM achieved up to 39% lower RMSE and an R2 of 0.990–0.994, outperforming advanced models. These results demonstrate a robust, generalizable solution for SCADA-based anomaly detection and early, interpretable fault localization across diverse wind turbine fleets.