Real-time rotor effective wind speed estimation using Gaussian process regression and Kalman filtering

Wai Hou Lio*, Ang Li, Fanzhong Meng

*Corresponding author for this work

    Research output: Contribution to journalJournal articleResearchpeer-review


    The use of state estimation technique offers a means of inferring the rotor effective wind speed based upon solely standard measurements of wind turbines. For the ease of design and computational concerns, typical wind speed estimators rely on a pre-computed mapping that describes the relationship from tip-speed ratio and pitch angle to the power coefficient. Typically, this mapping is built using numerical simulations under steady inflow conditions. Thus, the mapping built by traditional methods does not well capture the influence of other turbine dynamics and atmospheric variations, thus, inevitably resulting in poor performance of the wind speed estimator. Therefore, the paper presents a framework of rotor effective wind speed estimator design that obviates the need for a pre-computed power coefficient mapping. Specifically, the proposed method reconstructs the mapping using Gaussian process regression with a small set of real-time turbine measurement data. Subsequently, the wind speed estimator is built based upon the regression-based model and an extended Kalman filter, enabling optimal estimation from standard turbine measurements. The proposed method was evaluated in normal operation and down-regulation, against benchmark models obtained from an aero-elastic code. The estimation errors of the power coefficient and wind speed were significantly reduced by the regression-based approach.
    Original languageEnglish
    JournalRenewable Energy
    Pages (from-to)670-686
    Number of pages17
    Publication statusPublished - 2021


    • Rotor effective wind speed
    • Kalman filtering
    • Gaussian process regression
    • State estimation
    • Down-regulation
    • Control


    Dive into the research topics of 'Real-time rotor effective wind speed estimation using Gaussian process regression and Kalman filtering'. Together they form a unique fingerprint.

    Cite this