Wind turbine inflow estimation using blade loads and scada: steps towards real-world viability

As offshore wind farms grow, they increasingly face operational difficulties caused by wake interactions. Turbines operating within wakes have a reduced energy yield and experience additional fatigue due to elevated turbulence. These effects can be mitigated through wind farm flow control – the coordinated operation of turbines towards optimising the flow within the farm. Many such control solutions require detailed flow information at each turbine.
We have recently published a novel wind field estimation methodology [1], where features extracted from blade loads and SCADA are used in a localised linear regression model to reconstruct instantaneous wind “snapshot” across the rotor plane. The reported implementation performed well, but in some circumstances suffered from underestimated wind speeds across the hub-area. Training and testing were both undertaken using medium-fidelity wind field and turbine representations (Mann turbulence, parametric wake models, BEM-based load response). Concerning practical applications, a natural question therefore arises regarding the viability of medium-fidelity training datasets for prediction on high-fidelity or real-world cases. More specifically, it would be highly beneficial if the required large training datasets could be generated using less computationally expensive methods. The current work tackles these various factors in three stages: 1) an enhanced methodology is presented which overcomes hub-area underprediction issues 2) a cross-fidelity implementation is attempted, from medium-fidelity training to high-fidelity prediction 3) further analyses are undertaken to better understand the characteristics of a minimally viable training dataset.
The original methodology relies strongly on blade-load signals, which are driven mostly by wind conditions towards the outer blade sections. This factor, combined with training data which often includes wake impingement across the turbine hub, results in a commonly underestimated wind speed at the “snapshot” centre. It was therefore hypothesised that a new feature related to the local wind speed across the rotor hub area might overcome these issues. To achieve this, the original set of features (blade root bending moments, pitch angles and rotational speed) were extended to include a smoothed nacelle-mounted anemometer signal. Through model based testing this additional feature was shown to overcome the deficit issue and reduce the wind sensing error, indicating that the proposed hypothesis was correct.
In the cross-fidelity implementation, the medium-fidelity-trained model made predictions using high-fidelity turbine response inputs, obtained by combining Large Eddy Simulations (LES) with an Actuator Line Method (ALM). The resulting wind field estimates showed significant inaccuracies, indicating that a model trained on medium-fidelity data can’t straightforwardly be used for high-fidelity predictions. Two key differences between these cases are the turbulent structure and the aero-elastic solver (BEM vs ALM). To better understand the observed limitations in cross-fidelity predictions, further investigations are being undertaken in which we assess how variations in turbulence structure alone impact the model’s predictive ability. This will allow us to begin identifying a minimally viable training dataset towards real-world implementations of the enhanced wind sensing technique.
[1] Fojcik, P., Hart, E., and Hedevang, E.: Wind turbine wake detection and characterisation utilising blade loads and SCADA data: a generalised approach, Wind Energ. Sci., 10, 1943–1962, https://doi.org/10.5194/wes-10-1943-2025, 2025.