Hybrid machine learning model for prediction of vertical deflection of composite bridges

Abstract

A novel hybrid model, based on machine learning technique, for quick and accurate prediction of the vertical deflection of steel–concrete composite bridges was developed. The model is a combination of a bagging (B) ensemble and an instance-based k-nearest neighbours (IBk), hence called the B-IBk. In the models, five easily determined input parameters (cross-sectional shape, concrete beam length, age of the bridge, height of main girder and distance between main girders) are used to obtain the output parameter (maximum vertical deflection). To develop the models, direct measurement data from 83 steel–concrete composite bridges located at different places in Vietnam were collected and used as input and output parameters. Standard statistical evaluation indicators (mean absolute error, correlation coefficient (R) and root mean square error) were used to validate and compare the models’ performance. The results showed that the performance of the novel hybrid model (B-IBk) for predicting the maximum vertical deflection (Y) of steel–concrete composite bridges was very good (R = 0.908) and better than that of the single IBk model (R = 0.875) on the testing dataset. The developed novel model is thus a promising tool for accurate prediction of the Y of steel–concrete composite bridges.