Research on the Prediction of Rigid Frame-Continuous Girder Bridge Deflection Using BP and RBF Neural Networks

Abstract

To solve the problem of excessive deflection in the post-operation process of a rigid frame-continuous girder bridge and provide a basis for the setting of its initial camber, this paper, based on the results of finite element analysis, uses three methods to predict and verify the deflection of a rigid frame-continuous girder bridge. The results show that the average deflection method can be used to fit the average deflection value for a relatively long period of time and predict the average deflection value for the next longer period of time. Both the back-propagation (BP) neural network model and the radial basis function (RBF) neural network model can predict deflection well, but the RBF neural network model has higher prediction accuracy, with a mean absolute error (MAE) of 2.55 cmm and a relative error not exceeding 1%. The prediction model established by the RBF neural network has higher stability, better generalization ability, and better overall prediction performance. The established model has some reference significance for similar engineering projects and can achieve the optimization of structural parameters.