Three-dimensional spatiotemporal wind field reconstruction based on physics-informed deep learning

In this work, a physics-informed deep learning model is developed to achieve the reconstruction of the three-dimensional (3-D) spatiotemporal wind field in front of a wind turbine, by combining the 3-D Navier–Stokes equations and the scanning LIDAR measurements. To the best of the authors’ knowledge, this is for the first time that the full 3-D spatiotemporal wind field reconstruction is achieved based on real-time measurements and flow physics. The proposed method is evaluated using high-fidelity large eddy simulations. The results show that the wind vector field in the whole 3-D domain is predicted very accurately based on only scalar line-of-sight LIDAR measurements at sparse locations. Specifically, at the baseline case, the prediction errors for the streamwise, spanwise and vertical velocity fields are 0.263 m/s, 0.397 m/s and 0.361 m/s, respectively. The prediction errors for the horizontal and vertical direction fields are 2.84° and 2.58° which are important in tackling yaw misalignment and turbine tilt control, respectively. Further analysis shows that the 3-D wind features are captured clearly, including the evolutions of flow structures, the wind shear in vertical direction, the blade-level speed variations due to turbine rotation, and the speed variations modulated by the turbulent wind. Also, the developed model achieves short-term wind forecasting without the commonly-used Taylor’s frozen turbulence hypothesis. Furthermore it is very useful in advancing other wind energy research fields e.g. wind turbine control & monitoring, power forecasting, and resource assessments because the 3-D spatiotemporal information is important for them but not available with current sensor and prediction technologies.