Digital twin of wind farms via physics-informed deep learning

The spatiotemporal flow field in a wind farm determines the wind turbines' energy production and structural fatigue. However, it is not obtainable by the current measurement, modeling, and prediction tools in wind industry. Here we propose a novel data and knowledge fusion approach to create the first digital twin for onshore/offshore wind farm flow system, which can predict the in situ spatiotemporal wind field covering the entire wind farm. The digital twin is developed by integrating the Lidar measurements, the Navier-Stokes equations, and the turbine modeling using actuator disk method, via physics-informed neural networks. The design enables the seamless integration of Lidar measurements and turbine operating data for real-time flow characterization, and the fusion of flow physics for retrieving unmeasured wind field information. It thus addresses the limitations of existing wind prediction approaches based on supervised machine learning, which cannot achieve such prediction because the training targets are not available. Case studies of a wind farm under typical operating scenarios (i.e. a greedy case, a wake-steering case, and a partially-operating case) are carried out using high-fidelity numerical experiments, and the results show that the developed digital twin achieves very accurate mirroring of the physical wind farm, capturing detailed flow features such as wake interaction and wake meandering. The prediction error for the flow fields, on average, is just 4.7% of the value range. With the accurate flow field information predicted, the digital twin is expected to enable brand new research across wind farm lifecycle including monitoring, control, and load assessment.