Open AccessPrediction of shear wave velocity in underground layers using Particle Swarm Optimization
Author(s) -
Mark Ruben Anak Upom,
Mohd Nur Asmawisham Alel,
Mariyana Aida Ab. Kadir,
Ali Yuzir
Publication year - 2019
Publication title -
iop conference series. materials science and engineering
Language(s) - English
Resource type - Journals
eISSN - 1757-899X
pISSN - 1757-8981
DOI - 10.1088/1757-899x/527/1/012012
Subject(s) - particle swarm optimization , mean squared error , mean absolute percentage error , artificial neural network , support vector machine , mathematics , coefficient of determination , ensemble forecasting , linear regression , particle velocity , algorithm , statistics , artificial intelligence , computer science , physics , acoustics
Shear wave velocity (V s ) is considered a key soil parameter in the field of earthquake engineering. The time-averaged shear wave velocity in the upper 30 m (V s30 ) layer of soil is used to classify seismic site class. In-situ V s test is sometimes unsuitable to the project’s need due to financial reasons, noisy environment on site or simply the lack of expertise. This paper attempts to develop a global prediction model for V s using Standard Penetration Resistance (N spt ), depth ( z) and soil type ( s t) as the independent parameters. Two approaches to modelling would be taken; a multi-linear regression (MLR) model and an ensemble (EN-PSO) model. The EN-PSO model attempts to improve upon the accuracy of the MLR model prediction ability using the ensemble learning method. A dataset was compiled from literatures for this paper. 5 Base models were developed: MLR, Random Forest (RFR), Support Vector Machine (SVR), Artificial Neural Network (ANN) and k-Nearest Neighbor (KNN) which are combined into an ensemble model named EN-PSO. The weights for EN-SPO was then calculated using Particle Swarm Optimization (PSO). The performance of each models were then compared and it was shown that EN-PSO was the best in terms of: MAE (Mean Absolute Error) = 22.085, MAPE (Mean Absolute Percentage Error) = 9.1 %, RMSE (Root Mean Square Error) = 31.741 and R 2 Coefficient of Determination) = 0.895. In addition, it was also shown that the EN-PSO model was able to improve upon the performance of the MLR model, which the most accurate among the Base models. Comparisons were also made between EN-PSO and other suggested Universal V s correlations and EN-PSO was shown to outperform the other correlation based on prediction using a modified Test set. Three new empirical correlations as alternative for the EN-PSO model was also presented.