Hybrid Data-Driven Machine Learning Framework for Determining Prestressed Concrete Losses

This study presents a hybrid data-driven machine learning framework utilizing Bayesian regularization-based artificial neural network (ANN) and genetic expression programming (GEP) to develop a robust prestressing loss mathematical model. The proposed framework utilizes a surveyed database of experimentally measured losses of 113 prestressed concrete girders. In addition, this study utilizes the experimental database to evaluate the prediction accuracy of the adopted procedures in design provisions (PCI and AASHTO LRFD). The study shows that the currently adopted procedures result in either inconsistent accuracy or an overestimation of the prestressed losses. On the other hand, the proposed model demonstrated higher accuracy and consistent prediction with respect to all variables. The proposed model resulted in an average predicted-to-measured ratio of 1.03 and a standard deviation of 0.20. The study provides a parametric study to quantify the contribution of each of the included variables that revealed that girder height h, area of prestressing reinforcement A(ps), and moment due to dead load M-g are the most influential parameters in the estimation. Furthermore, a statistical evaluation of the different prestress loss estimation methods utilizing Bayesian parameter estimation demonstrated the robustness of the proposed method compared to the existing approaches.