INVESTIGATING THE USABILITY OF PHYSICS INFORMED MACHINE LEARNING APPROACHES FOR WIND FARM PLANNING

Determining wind energy yield and predicting wind farm performance is challenging due to the uncertain behavior of wind and the need for multiple CFD simulations across various scenarios and parameters, such as tree height, roughness, and wind direction. These simulations are computationally expensive and time-consuming, making Monte-Carlo methods impractical for extensive site assessments. This work explores the feasibility of using two types of neural networks—data-driven and physics-informed—for wind resource assessment. One network learns from CFD simulation data, while the other is trained directly with the physical equations (RANS and Navier-Stokes). Both networks are configured as feed-forward convolutional neural networks, with a loss function that incorporates residuals from either the data or the physical equations. The L-BFGS optimization algorithm is employed to minimize this loss and determine the network hyperparameters. The network predictions are compared with CFD simulations for two cases: flow over complex terrain and flow over a hill. The accuracy and limitations of these networks are evaluated by examining the multi-objective loss function, which includes errors from both data and physical equations. The trained networks demonstrate promise in accurately simulating wind resource assessments and offer potential improvements in accelerating CFD setups.