Developing hybrid artificial neural network model for predicting uplift resistance of screw piles

Abstract

The pull-out capacity of screw piles is affected by the complex underground and geological conditions, the helical pile’s configuration, and the penetration depth. Several experimental, theoretical, and neural network methods are available to predict the pull-out capacity of these piles. However, the weaknesses of ANN with regard to slow rates of convergence as well as in finding reliable testing outputs with reasonable errors are known to be major drawbacks of implementing ANN-based techniques. The present study aimed to develop an ICA-ANN-based model to estimate the pull-out capacity of screw piles in a simple way. A total of 36 experimental observations were collected and used to train, test, and optimize the ANN using an imperialist competitive algorithm (ICA). The developed ICA-ANN model can be considered an effective method for predicting the ultimate pull-out resistance of the helical screw piles since excellent agreement is obtained with respect to the reliability of the proposed model. The overall (training and testing) errors obtained for the proposed ICA-ANN model in comparison with the experimental data are 0.706, 0.17, and 0.996 for the mean absolute error (MAE), root-mean-square error (RMSE), and correlation factor (CF) respectively.