An Attention-Based LSTM Neural Network for Condition Prediction of Large-Span Cantilever Casting PC Continuous Box-Girder Bridges under Construction

For large-span cantilever casting prestressed concrete (PC) continuous box-girder bridges, the actual structure condition during construction stage cannot be consistent with the theoretical structure condition. As a consequence, an attention-based Long Short-Term Memory (LSTM) neural network prediction model is proposed in this paper. First, a bridge condition prediction framework is established, which can consider geometric dimensions, material characteristics, loads and time-dependent parameters. Then, numerical models of two large-span bridges are established for the theoretical structure condition analysis. Next, based on the theoretical and actual data of the bridge structure, the attention-based LSTM neural network is used to dynamically forecast the expected bridge condition. Finally, by comparing with the grey system theory and Kalman filter algorithm, the results show that the method proposed in this paper could improve the accuracy and efficiency of bridge condition prediction.