Wind inflow observation from load harmonics via neural networks: A simulation and field study

In this work, feed-forward neural networks are used to estimate the inflow of a wind turbine in real time during operation. A previously published method achieved this capability by correlating the turbine response to the wind inflow via a piecewise-linear model. Specifically, the existing formulation used the once-per-revolution harmonics of both in- and out-of-plane blade bending moments to estimate four wind parameters: the vertical and horizontal shears, and the vertical and horizontal misalignment angles. The novel method presented here builds on the same concept, but replaces the linear model with neural networks. Compared to the previous approach, the neural-based one is simpler to tune and use, because it decouples the wind parameters, which can be observed each one independently from the others. The proposed formulation is tested using both aeroservoelastic simulations and field data obtained on a 3.5 MW machine. In the field, independent measurements of the inflow are provided by a nearby hub-tall met mast, providing data for the training and verification of the wind observer. Notwithstanding some limitations of the experimental setup, results indicate that the neural networks produce estimates of the shears and yaw misalignment angle with accuracies compatible with typical use cases in wind turbine and wind farm control. Additionally, the neural-based observer appears to slightly outperform the original formulation in the lower wind speed range, where stronger nonlinearities in the load-wind mapping can occur.