Enhancing load capacity prediction of column using eReLU-activated BPNN model

Abstract

In structural engineering, accurately predicting the load-carrying capacity of columns is paramount for ensuring the safety and efficiency of construction projects. This study introduces an innovative approach to tackle this crucial challenge through the development and evaluation of an Exponential Rectified Linear Unit (eReLU) activated Backpropagation Neural Network (BPNN) model. The methodology initiates with the compilation of a comprehensive dataset, consisting of 120 samples representing column load-carrying capacity. This dataset undergoes meticulous curation, integrating findings from literature research and parametric analysis. Importantly, prior to model training, the dataset is subjected to thorough preprocessing to address outliers and undergo normalization. The eReLU activated BPNN model emerges as a superior performer when compared to conventional ReLU activated BPNN models. Demonstrating excellence in predictive accuracy, this novel model outperforms its ReLU-based counterparts. Key performance metrics, including a Mean Absolute Error (MAE) of 197.486 kN, a Root Mean Squared Error (RMSE) of 252.711 kN, and an R-squared (R2) value of 0.957 in the test set, underscore the model's robustness. Additionally, a feature importance analysis employing SHAP (SHapley Additive exPlanations) analysis transcends mere variable identification, emphasizing the significance of key variables such as e/2b, fyofyi, and f′c in predicting column load-carrying capacity. The eReLU activated BPNN model's holistic consideration of variable correlations significantly enhances its predictive capabilities, distinguishing it from ReLU-based models. This study represents a substantial leap forward in load-carrying capacity prediction methodologies, offering a robust and efficient model with broad applications in structural engineering.