Transient temperature field prediction on anti-icing wing surface based on convolutional temporal networks

Electric heating for anti-icing is a crucial technique to prevent icing on aircraft wings. Accurate prediction of transient temperature fields on the anti-icing surface of a wing is essential for optimizing the design of electric heating systems. To achieve a quick prediction of these transient temperature fields and shorten the optimization cycle of electric anti-icing systems, we propose a predictive method that couples the proper orthogonal decomposition (POD) with the convolutional temporal networks (CTN). This method first utilizes POD to perform dimensionality reduction on the temperature data. Subsequently, with the operation parameters as input and the reduced modal time coefficients as output, we construct a convolutional temporal network based on the 1D convolutional neural networks (1DCNN) and the temporal convolutional networks (TCN), incorporating a multi-head attention mechanism (MHA) to highlight key features. We introduce the sample accuracy (S_A) evaluation metric to evaluate the accuracy of transient temperature field predictions. Additionally, we explore the impact of model hyperparameters on the prediction performance and validate the effectiveness of the network structure through ablation experiments. Experimental results demonstrate that the proposed method achieves a S_A of 94.4% on the test set, indicating a high accuracy in predicting transient temperature fields on anti-icing wing surfaces.