Support Vector Machine Based Hybrid Model for Prediction of Road Structures Construction Costs

International Scientific Conference "People, Buildings and Environment 2018"

生态环境科学

Petrusheva, S.^1 ; Car-Pui, D.^2 ; Zileska-Pancovska, V.^1

University Ss Cyril and Methodius, Faculty of Civil Engineering, Skopje, R. Macedonia

1000, Macedonia^1

University of Rijeka, Faculty of Civil Engineering, Rijeka

51 000, Croatia^2

关键词: Coefficient of determination;    Construction projects;    General regression neural network;    Mean absolute percentage error;    MLP (multilayer perceptron);    Process-based modeling;    Project sustainability;    Rbfnn(radial basis function neural network);   

Others  :  https://iopscience.iop.org/article/10.1088/1755-1315/222/1/012010/pdf
DOI  :  10.1088/1755-1315/222/1/012010

学科分类：环境科学（综合）

来源: IOP

【 摘 要 】

Cost prediction in early stages of construction projects is one of the crucial problems of project sustainability. Previous research has been aimed at process based and data driven model development by using various techniques, e.g. regression analysis, support vector machine (SVM), neural networks etc. According to the research results, neither of the techniques can be considered the best for all circumstances. Therefore, the research has been redirected towards hybrid modelling, i.e. combination of different techniques. In this research, for prediction of the target variable - real construction cost of road structures, available variables: contracted construction cost, contracted construction time and real construction time and cost, hybrid model - combination of SVM technique (data-driven model) and Bromilow time-cost model (TCM) (process-based model) have been used. Five hybrid models have been built for comparison purposes: SVM-Bromilow TCM, LR-Bromilow TCM, RBFNN-Bromilow TCM, MLP-Bromilow TCM and GRNN-Bromilow TCM, combining Bromilow TCM with SVM, LR (linear regression), RBFNN (radial basis function neural network), MLP (Multilayer perceptron) and GRNN (general regression neural network), respectively. The highest accuracy has been obtained with SVM-Bromilow TCM with mean absolute percentage error (MAPE) 1.01% and coefficient of determination (R2) 97.61%.