Presentations

Abstract

Anemometers, conventional measurement devices for wind speed on wind turbines, often yield imprecise results due to disturbances from rotor blades, inherent methodological errors, or insufficient calibration. However, precise wind speed measurements are advantageous as they can be used for power curve validation and enhancing wind turbine control algorithms, thereby increasing the power production and reducing the mechanical loads and wear. A variety of wind speed estimation techniques exist today with the goal of correcting and improving anemometer wind speed measurements. These include observers based on theoretical models, additional sensors, the temporary use of costly LiDAR (Light Detection And Ranging) measurements, and the formulation of transfer functions to correct the anemometer wind speed. This contribution begins with a comprehensive review on these methodologies. Expanding on the state-of-the-art, this study proposes an approach that employs artificial intelligence algorithms without the reliance on additional sensors. Instead, the wind speed estimation and correction algorithm will use readily available measurements such as blade bending moments, hub accelerations, pitch and yaw angles, power, and rotor speed to correct the anemometer. Various combinations of these measurements will be used as inputs to the algorithms to find the optimal approach. Regarding artificial intelligence algorithms, different types of neural networks and decision trees are implemented and compared based on criteria such as accuracy, resource usage, and computational speed. Previous studies have implemented similar approaches on simulated data. However, it has been shown that wind speed estimates derived from simulations may underperform in real-world conditions. To address this discrepancy, the present study employs measured data from at least one turbine in a wind farm and one further experimental wind turbine. Both the wind farm and the experimental wind turbine are equipped with accurate LiDARs and measurement masts, which will serve as the wind speed reference after data processing to account for errors such as wake effects from a turbine on the measurements. In addition to finding algorithms to accurately estimate the wind speeds, the power curve of the turbines can be validated. The use of measured data will furthermore allow for the examination of additional challenges present when deploying the algorithms in the field. These challenges include the highly variable wind speeds, sensor noise, high sampling rates used in the wind turbine control algorithms, the necessity for fast algorithm computational speed with limited resources, and the issue of managing large volumes of data for algorithm training. In addition to the measurement data, simulation models of the wind turbines investigated in this study are available and can be used to generate a separate training data set. Algorithms trained using simulated and measured data can then be compared. Overall, the goal is to find an accurate and fast algorithm to correct the anemometer wind speed measurement based on artificial intelligence.