• Proposes a novel PSO-GRNN framework integrated with enhanced NeRF for high-fidelity 3D reconstruction and dynamic visualization of transmission tower icing deformation. • Introduces transient encoding and progressive positional encoding into NeRF, significantly compressing per-scene training time while improving characterization of icing dynamics. • Designs an adaptive inertia weight-based multi-objective PSO algorithm to optimize GRNN’s smoothing factor and architecture, enhancing prediction accuracy and generalization capability. • Achieves state-of-the-art performance with a Chamfer Distance of 1.24 mm, mean absolute error below 0.89 mm, and a 30%+ improvement in rendering efficiency at 31.5 fps. • Provides a robust, automated solution for icing deformation monitoring that enhances early-warning capacity and offers insights applicable to other truss structure analyses. Transmission line icing disasters pose a serious threat to the safe and stable operation of power grids. Traditional monitoring methods, which rely on meteorological extrapolation or manual inspection, struggle to accurately quantify ice thickness distribution and the resulting complex spatial deformation. To address this, this study proposes an integrated three-dimensional visualization framework that combines high-fidelity modeling with intelligent deformation prediction. The framework features two key innovations: an enhanced modeling approach based on Neural Radiance Fields (NeRF) and a deformation prediction algorithm based on a Particle Swarm Optimization–Generalized Regression Neural Network (PSO-GRNN). It enables dynamic visualization of the entire process from icing observation to deformation appearance. Experimental results are as follows. The Chamfer distance reaches 1.24 mm - approximately 34% lower than that of the next best method, Neural Feature Fields (NFF). The mean absolute error in deformation prediction is 0.89 mm, which outperforms Temporal Convolutional Networks (TCN) at 1.24 mm and Long Short-Term Memory networks (LSTM) at 1.57 mm. The average rendering frame rate reaches 31.5 fps, representing a 30% improvement over the Point Sprite method. This work provides a high-precision and efficient 3D visualization solution for monitoring icing-induced deformation on transmission towers. By fusing implicit neural representations with explicit deformation parameters, the method enables real-time, synchronized rendering of ice growth and structural response while maintaining sub-millimeter-level accuracy. It enhances the level of automation and the early-warning capability in deformation monitoring. Furthermore, the proposed framework offers a reference technical pathway for deformation analysis of similar truss structures.
Pan et al. (Wed,) studied this question.