Hybrid neural network techniques for friction capacity prediction in concrete pile foundations

Abstract

The advancement of novel data mining and optimization algorithms has significantly enhanced traditional engineering structural analysis models, particularly those based on swarm intelligence. This study delves into refining the neural assessment of shaft friction capacity in driven pile systems by exploring the social behavior of four hybridized algorithms: Wind-Driven Optimization (WDO), Spotted Hyena Optimization (SHO), Grasshopper Optimization Algorithm (GOA), and Moth-Flame Optimization (MFO). Four crucial influencing variables - pile length (m), diameter (cm), effective vertical stress (Sv), and undrained shear strength (Su) - are considered in constructing the requisite dataset. After applying optimized structures, each ensemble undergoes a sensitivity analysis based on its individual swarm size. The predictive precision of the models is compared using the results of two sensitivity analyses. Neural network simulations exhibit improved results with an increased number of neurons in a single hidden layer. The root mean square errors (RMSEs) for the training and test datasets, employing Multilayer Perceptron (MLP)-based solutions, are (0.05241, 0.32861, 0.06155, and 0.03874) and (0.04334, 0.18155, 0.05382, and 0.03626), respectively. In the training and testing datasets for proposed predictive models using WDO, SHO, GOA, and MFO, R-2 values of (0.996, 0.853, 0.992, and 0.997) and (0.985, 0.732, 0.997, and 0.997) were found, respectively. Notably, MFO outperforms its counterparts when integrated with MLP for predicting engineering solutions.