- 期刊
- OpenAccess
Prediction of Pressure Coefficient for SACCON Wing using Artificial Neural Network
摘要
This study explores the application of artificial neural networks (ANN) for predicting the coefficient of pressure (CP), an essential parameter in aerodynamic studies, on the SACCON wing. Traditional methods such as computational fluid dynamics (CFD) and wind tunnel tests have several drawbacks, including high computational cost, time consumption, and inconsistent data. To address these issues, this study proposes a multi-layer perceptron (MLP) neural network as an alternative. The training data for the ANN is derived from previously conducted wind tunnel experiments on the SACCON wing. The network's structure includes three hidden layers and employs the Levenberg-Marquardt backpropagation algorithm for weight updates. Data pre-processing involves normalization, and the dataset is split into training, testing, and validation sets. The results indicate that the ANN can capture general trends in CP data. However, it struggles with detailed and irregular predictions specifically in predicting tip vortex and suction peak, leading to underfitting, partly due to a lack of sufficient training data at location where vortex interaction occurred. The optimal network configuration includes a learning rate and regularization parameter of 1 x 10^(-5) and hidden layers containing 10, 10, and 3 neurons, respectively. Despite achieving reasonable performance metrics (RMSE = 0.1555, R-value = 0.9855, R^2 = 0.9711), the network's predictions on new data are less accurate due to underfitting. Future improvements suggest increasing the training dataset size or enhancing feature engineering to improve the ANN's predictive capabilities. Additionally, using ANNs during the testing phase can reduce time, manpower, costs, and the number of data points required to define aerodynamic performance.