Inclination prediction of a giant open caisson during the sinking process using various machine learning algorithms

Abstract

Machine learning (ML) models based on 12 ML algorithms were established to predict the open caisson inclination. The prediction performance of these models was evaluated in the main pier open caisson monitoring of the Changtai Yangtze River Bridge Project, and the prediction performance was compared based on prediction results, corresponding residuals, prediction accuracy, calculation time and training sample dependence to determine the suitable prediction models. Then, the influence of 3 key parameters was analysed, and the models with the best prediction performance were optimized. The results showed that 7 ML algorithms were suitable for inclination prediction of this project. The best performance was obtained using models based on extra trees (XT) and k-nearest neighbour (KNN). The influence of the 3 key parameters on the prediction accuracy was determined, which is beneficial to further optimize the prediction models. After optimization, the root mean square error and the calculation time of the KNN model were decreased by 45% and 33%, and the best prediction accuracy was obtained using the default parameter settings of the XT model.