Traffic Load Simulation for Long-Span Bridges Using a Transformer Model Incorporating In-Lane Transverse Vehicle Movements

Abstract

Traffic load simulation (TLS) is critical for the design and assessment of long-span bridges. Traditional methods, such as Monte-Carlo sampling and Cellular Automaton, rely on actual traffic data for load generation and evolution. However, they often overlook in-lane transverse movements, which are vital for precise bridge component assessment. This paper presents a TLS framework that incorporates in-lane transverse movements for long-span bridges. We select eight parameters as input features for a Transformer-based deep learning model, designed to predict both longitudinal and transverse vehicle speeds. The TLS process begins with spatial-temporal traffic load monitoring on the target bridge. Monte-Carlo sampling generates vehicle data, and the trained Transformer model simulates traffic evolution. A case study on a 1490-meter main-span suspension bridge illustrates the proposed method. Traffic trajectories were captured using a multi-vision system and reconstructed to minimize errors. The Transformer model was trained with optimized hyperparameters, enabling the completion of TLS on the entire bridge deck. We also compare the performance of other deep learning models, evaluate the accuracy of transverse distribution in TLS, and discuss its potential applications in future bridge assessments. The proposed TLS method enhances current practices by accurately simulating transverse vehicle positions on bridge decks, thereby improving the fidelity of microscopic traffic simulations and enabling more precise fatigue damage assessments of bridge components.