A physics-informed neural network for turbulent wake simulations behind wind turbines

Fast simulations of wind turbine wakes are crucial during the design phase of optimal wind farm layouts. Wind turbine wakes affect the performance of downstream turbines. Physics-informed neural networks (PINNs), a deep learning approach to simulate dynamical systems governed by partial differential equations, are gaining traction in computational fluid dynamics due to their fast inference capability. We developed a PINN model using the 2-equation k−ε model and the actuator disk method to simulate the wakes behind the wind turbines. Crucially, training of the developed PINN model does not rely on high-fidelity simulation data, thus reducing the end-to-end training time by saving simulation data generation time. We tested the model against traditional solvers and field data to simulate the turbulent wake behind the HOLEC WPS 30/3 Wind Turbine from Sexbierum and a three-blade 630-kW Nibe-B wind turbine. Detailed computational studies are completed to establish convergence properties with increasing sampling collocation points and the number of graphical processing units. A transfer learning strategy is introduced to accelerate training for new scenarios resulting in a 5× speedup. Our results establish the efficacy of the PINN model in simulating turbulent flows. Compared to field data, our PINN model and traditional Reynolds-averaged Navier–Stokes (RANS) numerical solvers, such as the shear stress transport k −ω and Reynolds stress model have similar errors, suggesting its utility as a replacement to these RANS solvers. The model architecture, trained weights, and code are available in https://github.com/quest-lab-iisc/PINN_WakeTurbulenceModel.