Posters

Abstract

Yaw misalignment constitutes a substantial performance deficit in wind turbine operation, driving both Annual Energy Production (AEP) losses and accelerated component fatigue. This study presents a real-world case that evaluates the accuracy of two distinct data-driven detection methodologies applied to a wind farm situated in the semi-arid region of Bahia, Brazil.  The investigation benchmarks a standard parametric framework against a novel non-parametric machine learning counterpart. The first method, based on the OpenOA library [1], utilizes a quantile-filtering mechanism followed by a cosine-based curve fitting for the power-yaw relationship. In contrast, the second method [2] implements a multi-stage pipeline designed for high-noise environments. This proposed framework integrates the Local Outlier Factor (LOF) algorithm to effectively isolate healthy operational data. Subsequently, it employs a Random Forest regressor to reconstruct the power curve without assuming a fixed cosine low relationship.  Preliminary benchmarking using one year of historical data reveals substantial divergence between the two frameworks, primarily driven by sampling resolution and temporal filtering by daytime and nighttime. Results indicate that both frameworks are sensitive to data resolution: one turbine exhibited opposing trends, with [1] estimates decreasing from -4.4° to -33.8° while the [2] demonstrated sensitivity by flipping from -8° in 1-minute sampled data to 12° in 10-minute sampled data. Such inconsistency suggests that the 10-minute average may suppress dynamic features present at high frequency or, conversely, that the 1-minute sampling may incorporate unfiltered measurement noise. On the other hand, the framework [1] also produced unreasonable incongruence in other cases due to temporal filtering, generating diurnal variations and shifting in one instance from -4.4° (Day) to -30.9° (Night), whereas the [2] framework, for example, in another turbine, indicates 24° (Day) to 4° (Night). The site is characterized by isolated terrain ruggedness, and pronounced diurnal thermal variance. These complex flow conditions introduce high wind shear, veer, and variable turbulence intensities, which are known to decouple the nacelle-measured wind direction from the true freestream wind vector, potentially biasing SCADA-based estimations, supporting the differences in results between day and night datasets [3]. Consequently, these environmental factors and resolution constraints limit the reliability of purely data-driven algorithms when applied in isolation. To mitigate this uncertainty and validate the SCADA-derived estimates, a field campaign will be carried out utilizing a ZX TM Nacelle-Mounted LiDAR (NML) measurements as ground truth and will be included. The study further correlates these frameworks' predictions with the absolute wind vectors captured by the NML campaign. This paper contributes to understanding the validity of SCADA-based frameworks to estimate yaw misalignment. [1] J. Perr-Sauer, et al. OpenOA: An Open-Source Codebase For Operational Analysis of Wind Farms. Journal of Open Source Software, 2022.  [2] F. R. de Simoni. Yaw misalignment analysis and correction algorithm: A data-driven framework. Instituto Federal de Educação, Ciência e Tecnologia de Santa Catarina, 2024. [3] L. Gao, et al. Catch the wind: Optimizing wind turbine power generation by addressing wind veer effects. PNAS Nexus, 2024.