The KAN-MHA model: A novel physical knowledge based multi-source data-driven adaptive method for airfoil flow field prediction

Abstract

The power generation efficiency and operational safety of wind turbines depend heavily on the aerodynamic performance of their airfoils, which is primarily controlled by the flow field. Traditional methods for predicting this flow field rely on solving the Navier-Stokes (NS) equations, which are computationally expensive and inefficient. To address these challenges, this paper proposes a novel neural network model, the KAN-MHA, which integrates a Kolmogorov-Arnold Network (KAN) with a Multi-Head Attention (MHA) mechanism, leveraging multi-source datasets that include wind tunnel experiments, XFoil results, and CFD simulations. By incorporating physical constraints, the KAN-MHA model achieves precise predictions of flow fields with a simpler architecture and reduced computational cost. Experimental results indicate that, compared to traditional multilayer perceptron (MLP) networks, the KAN-MHA model achieves a substantial reduction in testing loss, while maintaining a significantly simpler architecture with fewer neurons. Through an analysis of the attention weights in the MHA mechanism, we found that MHA effectively guides the network to focus on regions with more intricate flow variations, thereby enhancing the model's ability to capture subtle flow field features with higher accuracy. As a result, the model demonstrates excellent prediction performance for aerodynamic coefficients and effectively identifies stall behavior in airfoils at high angles of attack. This work provides a novel approach for the design and optimization of wind turbine airfoils, offering valuable insights for enhancing aerodynamic performance under complex flow conditions.