Spatiotemporal wind field prediction based on physics-informed deep learning and LIDAR measurements

Spatiotemporal wind field information is of great interest in wind industry e.g. for wind resource assessment and wind turbine/farm monitoring & control. However, its measurement is not feasible because only sparse point measurements are available with the current sensor technology such as LIDAR. This work fills the gap by developing a method that can achieve spatiotemporal wind field predictions by combining LIDAR measurements and flow physics. Specifically, a deep neural network is constructed and the Navier–Stokes equations, which provide a good description of atmospheric flows, are incorporated in the deep neural network by employing the physics-informed deep learning technique. The training of this physics-incorporated deep learning model only requires the sparse LIDAR measurement data while the spatiotemporal wind field in the whole domain (which cannot be measured) can be predicted after training. This study, which can discover complex wind patterns that do not present in the training dataset, is totally distinct from previous machine learning based wind prediction studies which treat machine learning models as “black-box” and require the corresponding input and target values to learn complex relations. The numerical results on the prediction of the wind field in front of a wind turbine show that the proposed method predicts the spatiotemporal flow velocity (including both downwind and crosswind components) in the whole domain very well for a wide range of scenarios (including various measurement noises, resolutions, LIDAR look directions, and turbulence levels), which is promising given that only line-of-sight wind speed measurements at sparse locations are used.