Uncertainty-aware anomaly detection for wind turbines using a physics-informed spatiotemporal feature modelling model

Duan, H, Zhang, Q ORCID: 0000-0002-0651-469X, Ou, H, Li, C, Zhang, W and Xu, Z (2026) Uncertainty-aware anomaly detection for wind turbines using a physics-informed spatiotemporal feature modelling model. Applied Energy, 426. ISSN 0306-2619

Preview

Text Uncertainty-aware anomaly detection for wind turbines using a physics-informed spatiotemporal feature modelling model.pdf - Published Version Available under License Creative Commons Attribution. Download (17MB) | Preview

Abstract

Wind turbines are prone to failures in critical subsystems such as the drivetrain and generator, making reliable early anomaly detection essential for safe and stable operation. With the increasing availability of monitored time-series data, data driven methods have emerged as an effective basis for anomaly detection and have demonstrated considerable potential for improving the reliability and safety of wind turbines. Nevertheless, conventional data driven approaches often struggle to adequately characterize the temporal dependencies in wind turbine operational data, the correlations among monitored variables in the feature space, and the physical constraints governing system operation, resulting in limited detection accuracy and robustness under varying operating conditions. To address these limitations, this paper proposes a physics-informed deep learning framework based on multi-sensor data, which integrates spatiotemporal feature modeling with physical consistency to improve the reliability of early anomaly detection and support prognostic functionality for wind turbines. The proposed framework is validated using a real SCADA dataset collected from two wind farms. Experimental results demonstrate that the proposed PITG achieves the best predictive performance across four datasets, with an average RMSE of 0.1397 and an average R² of 97.21%. Meanwhile, the physics-informed embedding module effectively reduces prediction uncertainty, decreasing the maximum prediction variance by approximately 27.25% compared with the baseline model. Combined with the inverse-variance multi-sensor fusion strategy, the proposed framework achieves an F1-score of 0.566 for system-level anomaly detection, validating its effectiveness and robustness under complex operating conditions.

Item Type:	Article
Uncontrolled Keywords:	Anomaly detection; Physics-informed neural network; Prognostics and health management; Wind turbine; SCADA data; 4007 Control Engineering, Mechatronics and Robotics; 40 Engineering; Networking and Information Technology R&D (NITRD); 7 Affordable and Clean Energy; 09 Engineering; 14 Economics; Energy; 33 Built environment and design; 38 Economics; 40 Engineering
Subjects:	T Technology > TA Engineering (General). Civil engineering (General)
Divisions:	Engineering and Built Environment
Publisher:	Elsevier
Date of acceptance:	4 August 2026
Date of first compliant Open Access:	1 September 2026
Date Deposited:	01 Sep 2026 14:03
Last Modified:	01 Sep 2026 14:03
DOI or ID number:	10.1016/j.apenergy.2026.128639
URI:	https://researchonline.ljmu.ac.uk/id/eprint/29252

View Item