基于POD-神经网络的格尼襟翼压力分布预测
PREDICTION OF PRESSURE DISTRIBUTION ON GURNEY FLAPS BASED ON POD-NEURAL NETWORK MODEL
-
摘要: 针对加装格尼襟翼翼型气动特性数值仿真计算效率较低这一问题, 提出一种融合本征正交分解(proper orthogonal decomposition, POD)与神经网络的混合预测方法. 通过POD方法从压力分布数据中提取前i阶主导模态, 构建表征压力分布核心特征的低维空间, 再基于各阶模态对应的特征向量重构数据集, 并通过神经网络模型分别学习其非线性映射规律, 最终集成多尺度预测结果实现压力分布的高精度重构. 该方法提出一种基于POD与神经网络的联合预测方法, 有效解决了传统单一模型在翼型表面压力分布预测中局部流动细节捕捉能力不足问题. 以加装格尼襟翼的RAE2822翼型为研究对象, 基于高精度CFD仿真生成多工况压力分布数据, 经验证仿真结果与实验数据最大偏差低于6%. 实验结果表明, 该方法的预测准确率提升至91.11%, 对不同工况差异的捕捉能力显著增强, 相较于随机森林算法, 预测精度提高了18.83%. 此外, 该方法在计算效率方面也有明显提升. 本研究不仅提高了加装格尼襟翼翼型气动特性预测的精度与效率, 还为解决类似复杂气动问题的数值仿真提供了新的思路.Abstract: To address the challenge of low computational efficiency in the numerical simulation of aerodynamic characteristics of airfoils equipped with Gurney flaps, this study proposes a novel hybrid prediction method that integrates proper orthogonal decomposition (POD) with neural networks. Traditional aerodynamic prediction methods based solely on computational fluid dynamics (CFD) or single machine learning models often struggle with high computational costs and limited ability to capture detailed local flow features, especially under varying flow conditions. The proposed approach first applies the POD method to extract the leading i principal modes from high-fidelity pressure distribution data, thereby constructing a compact low-dimensional feature space that encapsulates the dominant flow structures around the airfoil. Each sample is then reconstructed using the corresponding modal coefficients, and multiple neural networks are independently trained to learn the nonlinear mapping relationships between flow conditions and these coefficients. This multi-model framework enables the prediction of pressure distributions with both high resolution and improved generalization across different flow regimes. The RAE2822 airfoil with an installed Gurney flap is selected as the test case, and pressure distribution datasets under multiple flow conditions are generated through high-precision CFD simulations. Validation against experimental data shows a maximum deviation of less than 6%, demonstrating the accuracy of the simulation process. The experimental results further indicate that the proposed hybrid method achieves a prediction accuracy of 91.11%, significantly outperforming traditional algorithms such as random forest by a margin of 18.83%. In addition to improved accuracy, the method also achieves notable gains in computational efficiency due to the reduced dimensionality and parallel training architecture. Overall, this research not only enhances the prediction performance for airfoils with Gurney flaps but also provides a generalized and scalable framework for accelerating aerodynamic simulations involving complex geometries and flow conditions. It offers valuable insights into integrating data-driven techniques with physics-based methods for advanced aerodynamic analysis and design optimization.