Abstract: To achieve high-precision prediction of bridge-coupled extreme stresses, the wavelet multi-resolution analysis method is adopted to decouple the coupled extreme stresses. The decoupled low-frequency data is taken as the trend item information, where the high-frequency data is considered as the vehicle load effect information. The trend item, after subtracting its mean, is the temperature load effect information. A bivariate Bayesian dynamic linear trend model (BDLTM), which introduces a time-varying trend term, is built to predict and analyze low-frequency extreme stress. GRU neural network model is provided to predict and analyze high-frequency extreme stresses. The dynamic coupled extreme stresses are predicted. The monitoring coupled data from Tianjin Fumin Bridge is provided to illustrate the feasibility and application of the proposed BDLTM-GRU model, the accuracy of which is compared with the single BDLTM model and single GRU model for verifying the high precision of the BDLTM-GRU model.