Prediction of shear wall residential beam height based on machine learning

Abstract

The beam height is an important design parameter that influences structural properties such as load-bearing capacity and stability of beams. In the early stages of structural design, the existing methods for determining beam height mainly include empirical formulae. However, empirical methods are highly subjective, lack accuracy, and are poorly adapted to complex conditions. This paper establishes a beam height prediction model for shear wall residential structures. Using structural design data from projects built by a real estate company across various regions in China, a large dataset of beam heights was collected. The Permutation Feature Importance (PFI) method and six unique machine learning (ML) models were used to rank the importance of input variables. The Gradient Boosting (GB) model, consistent with the feature ranking obtained from PFI, was selected. The model evaluation method was then used to select the number of input features for the GB model, and grid search and K-fold cross-validation were employed to optimize the GB prediction model. This model was compared with a prediction model obtained from a Back Propagation Neural Network (BPNN). Finally, the SHAP method was used to interpret the "black box" machine learning model. The results show that the GB model has higher accuracy compared to the BPNN model, and the input features of the proposed GB model contribute to the beam height in accordance with mechanical laws, demonstrating the model's rationality. The research findings can provide a reference for initial beam height design.