Machine-Learning-Based Surrogate Modeling of Aerodynamic Flow Around Distributed Structures

Abstract

A machine-learning-based surrogate modeling method for distributed fluid systems is proposed in this paper, where a dimensionality reduction technique is used to reduce the flowfield dimension and a regression model is used to predict the reduced coefficients from the input parameters. The surrogate modeling method is specifically designed to tackle the fluid systems involving distributed aerodynamic structures, and its performance is illustrated by the application on the wake flow around wind turbine arrays in an atmospheric boundary layer. The main idea is to first decompose the whole fluid domain into subdomains, then carry out surrogate modeling for each subdomain by treating both the boundary information and the distributed flow parameters as the input parameters, and finally obtain the whole flowfield by combining the flowfield of each subdomain with the consideration of the matching condition at the subdomain interface. The proposed surrogate modeling method is applied to two test cases: a one-dimensional Poisson equation and a high-fidelity wind farm wake model. The results demonstrate the great efficiency and accuracy of the surrogate model and its excellent scalability to distributed systems of different scales.