Deep learning-based minute-scale digital prediction model for temperature induced deflection of a multi-tower double-layer steel truss bridge.

This article focuses on developing and evaluating a deep learning-based digital prediction model for temperature-induced deflection in a long-span multi-tower double-layer steel truss bridge. Using structural health monitoring (SHM) data from the Beikou Bridge in China, the study compares three neural network architectures—long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM), and a Transformer variant—under two learning strategies differing in input data composition. Results indicate that incorporating both historical temperature and deflection data (Strategy B) significantly improves prediction accuracy over using temperature data alone (Strategy A). Among the models tested, the Transformer-variant network, leveraging its self-attention mechanism, demonstrates superior performance in capturing key temporal features and providing stable, accurate predictions of temperature-induced bridge deflection.