Posters

Abstract

Assessing wind energy yield and predicting wind farm performance is challenging due to the unpredictable nature of wind and the need for numerous computational fluid dynamics (CFD) simulations across varying scenarios and parameters, such as tree height, surface roughness, and wind direction. These simulations are computationally intensive and time-consuming, making Monte Carlo methods impractical for comprehensive site evaluations. This study investigates the potential of two types of neural networks—data-driven and physics-informed—for wind resource assessment. The data-driven network is trained on CFD simulation data, while the physics-informed network learns directly from physical equations, including the Reynolds-Averaged Navier-Stokes (RANS) and Navier-Stokes equations. Both networks are designed as feed-forward convolutional neural networks with loss functions that incorporate residuals from either data or physical equations. The L-BFGS optimization algorithm is used to minimize this loss and tune the network hyperparameters. Network predictions are benchmarked against CFD simulations in two scenarios: flow over complex terrain and flow over a hill. The performance and limitations of these networks are evaluated by analyzing the multi-objective loss function, which captures errors from both data and physical equations. The results show that the trained networks can effectively simulate wind resource assessments, offering significant potential to speed up CFD setup processes.