Wind Turbine Hybrid Physics-Based Deep Learning Model for a Health Monitoring Approach Considering Provision of Ancillary Services

Abstract

Assessing the overall condition of wind turbines (WTs) in operation is challenging due to their intricate nature. This becomes even more complicated when WTs provide ancillary services and respond to grid requirements under curtailment modes. Multiple models are required to effectively evaluate the WTs' healthy condition, which can be unmanageable and impractical, particularly for large-scale wind farms. This article proposes a novel hybrid physics-based deep learning framework to accurately approximate the time-varying correlation between control sequences and system response, reflecting the aerodynamic nonlinearity of the 5-MW offshore WT model, designed and tested by the National Renewable Energy Laboratory (NREL). Another layer of this study's novelty relies on proposing a computationally efficient weakly supervised method that uses the hybrid structure to detect degradations and anomalies considering curtailment operation. Then, a self-learning classification approach is employed to iteratively update the best-tuned classifier, dynamically learning unforeseen abnormalities from brand-new anomalies during active operations. The proposed anomaly detection strategy deals with system uncertainties, such as wind stochasticity, power curve variations, and different sparsity levels in the datasets. The results of the proposed approach show promise in improving health monitoring performance, leading to a more efficient and accurate assessment of the overall condition of WTs.