Open AccessCFD-based surrogate modelling of urban wind farms using artificial neural networks: double rotor arrangements
Author(s) -
Sadra Sahebzadeh,
Abdolrahim Rezaeiha,
Hamid Montazeri
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/2042/1/012093
Subject(s) - artificial neural network , computational fluid dynamics , mean squared error , rotor (electric) , surrogate model , turbine , power (physics) , computer science , obstacle , wind power , wind speed , simulation , control theory (sociology) , marine engineering , engineering , artificial intelligence , mathematics , machine learning , statistics , mechanical engineering , meteorology , aerospace engineering , physics , electrical engineering , control (management) , quantum mechanics , law , political science
Extensive characterization studies are required to identify optimal wind farm layouts and achieve high power density, i.e., power per land area. Performing such studies using experimental or high-fidelity numerical methods can be timely and computationally expensive. To alleviate this obstacle, surrogate models can be developed to mimic the behavior of the simulation/experiment. In this paper, a shallow feed-forward artificial neural network (ANN) surrogate model is developed. The Levenberg-Marquardt algorithm is used to train a model with 3 layers and 10 hidden nodes. The model correlates the arrangement of a double-rotor vertical axis wind turbine array, as the fundamental generating cell of the wind farm, with its overall power performance. The inputs are the relative distance (R) and angle (®) between the rotors, and the output is the overall power coefficient of the array. In total, 96 CFD-simulated arrangements are used as data points to train, validate and test the model. The trained model has a mean square error of 2.10 × 10 -5 and R-squared of 0.99, indicating its accuracy and generalizability. The average and maximum errors are 3% and 10%, respectively. The employed method can be expanded to accommodate more rotors towards optimal urban wind farm layout design.