A data-driven surrogate model framework based on CFD simulations to accelerate wind energy yield assessment

Alessandro Gambale S.K. Raghunathan Srikumar G. Mosca Tsionas, Ioannis [UCL] Llaguno-Munitxa, Maider [UCL] Stephan, André [UCL]

A data-driven surrogate model is developed using deep neural networks, trained from a set of CFD simulations to predict velocities on a 3D grid for a specific inlet velocity condition. The objective is to considerably reduce the number of CFD simulations to be performed when assessing the annual wind energy yield of rooftop micro wind turbines. Steady, incompressible RANS simulations are performed to obtain the wind velocities in the Northern District region of Brussels, Belgium using a modified k-𝜔 SST turbulence model implemented with an improved Atmospheric Boundary Layer approach. Velocity fields from simulations serve as the labels to train the surrogate model, and their corresponding inlet velocity conditions are the features. Prediction of the 3D velocity field is then performed for a new inlet velocity,and the performance of the trained model is assessed by comparing the results with the ones obtained from the CFD simulations. Based on the obtained velocity fields, rooftop locations with high velocities are chosen for the wind turbine installation, and the annual energy yield is calculated at each of these locations. Finally, the speedup in computational time obtained by utilizing a surrogate model in conjunction with CFD for a complete assessment of annual wind energy yield is discussed.