Enhancing Wind Field Prediction and Reconstruction around Windbreak Walls along High-Speed Railway by Advanced Neural Network Architectures: Accuracy and Stability Assessment

Abstract

Accurate prediction of wind fields around high-speed railway (HSR) infrastructure is critical for operational safety and energy efficiency. This study evaluates neural network approaches for predicting wind fields around HSR windbreak walls, focusing on transformer models. Field measurements were conducted using 15 anemometer masts arranged inside and outside windbreak walls on the Lanzhou–Xinjiang railway. We compared multiple deep learning architectures (multilayer perceptron, long-short-term memory, temporal convolutional network and transformer) for predicting interior wind conditions based on exterior measurements. The key findings are summarized as follows: (1) prediction accuracy improved substantially with longer historical contexts (10–60 timesteps); (2) significant spatial variability exists in wind predictability across measurement locations; (3) feature importance analysis identified critical measurement points, enabling cost-effective maintenance strategies and optimized sensor deployment; and (4) sequence mean filling performed best among the three tested strategies for handling missing data, maintaining positive predictive power even with substantial sensor loss. Among these models, the transformer model achieved the best overall performance (R² = 0.9665 at T = 60), with its advantage becoming most pronounced at longer historical contexts. These findings have important implications for railway safety and wind energy applications, enabling more efficient monitoring networks and robust forecasting systems. The demonstrated effectiveness of transformer models represent a significant advancement in applying attention-based architectures to infrastructure monitoring challenges.