Probabilistic Prediction of Ice-Shedding Jump Height of Overhead Transmission Lines Using a Dimensionless-Group-Guided Bayesian Neural Network

The jump height of overhead transmission lines after ice shedding is closely related to electrical clearance, and structural safety. For different conductor parameters, span lengths, horizontal stress, and icing conditions, transient finite-element analysis can provide detailed dynamic responses, but repeated simulations are not convenient for fast engineering assessment. In addition, a deterministic prediction model only gives a single jump-height value, and the reliability of this value is difficult to judge when the input condition is close to the boundary of the sampled range. In this study, a Dimensionless-Group-Guided Bayesian Neural Network (DG-BNN) is developed to predict the ice-shedding jump height and estimate the associated uncertainty. The original physical variables are first transformed into seven dimensionless Pi-groups according to Buckingham Pi dimensional analysis. These variables describe the main effects of geometry, mass distribution, stress state, and ice shedding in a compact form. The network is trained through a two-stage procedure. A heteroscedastic regression model is first obtained with a trend-regularization term based on a fixed second-order polynomial, and then the deterministic layers are converted into Bayesian layers for probabilistic inference. Monte Carlo sampling is used to calculate the predictive mean, epistemic uncertainty, and aleatoric uncertainty. Temperature scaling is further introduced to adjust the prediction intervals on the validation set. The comparison with several regression models shows that DG-BNN can maintain accurate jump-height prediction while giving calibrated uncertainty information. On the FEM-generated test set, DG-BNN achieved an R2 of 0.986, an RMSE of 0.958 m, and an MAE of 0.649 m. After temperature calibration, the coverage probability of the 90% prediction interval reached 89.96%. The model can therefore serve as a fast surrogate tool for ice-shedding response assessment, especially when the reliability of the predicted result needs to be considered.