Study on the prediction of aerodynamic coefficients for crescent-shaped ice-accreted conductors using a BP neural network

This paper presents a nonlinear mapping prediction model for aerodynamic coefficients of ice-accreted conductors, which is established using a BP neural network. The model integrates experimental data and computational fluid dynamics numerical simulation data from multiple sources pertaining to crescent-shaped ice-accreted conductors. A dual-hidden-layer network architecture was developed, utilizing wind speed, ice thickness, conductor diameter, and wind attack angle as input variables, with lift coefficient, drag coefficient, and torque coefficient serving as the output variables. A fully connected three-layer architecture was established via the Keras sequential model, enabling the effective capture of high-order nonlinear relationships within the aerodynamic characteristics of ice-accreted conductors. The employment of Kernel SHAP analytics to assess neural network hidden layer feature interactions ensures the interpretability of the prediction results. The results show that when the model is used for prediction, the coefficients of determination of lift coefficient, drag coefficient, and torque coefficient are as high as 0.912, 0.970, and 0.964, respectively, and the mean square error is as low as 0.0300, 0.0095, and 0.0081, respectively. Standardized residual analysis reveals that in the test set, the residuals for lift coefficient, drag coefficient, and torque coefficient are predominantly in the range of −0.2 to 0.2, with rates of 85.92%, 96.24%, and 97.65%, respectively. These results suggest that the model exhibits outstanding generalization capabilities and can be extensively utilized for predicting aerodynamic coefficients following ice accretion on various conductor types. In addition, the model offers substantial auxiliary benefits for forecasting galloping and evaluating risks in transmission lines.