Development of an artificial neural network model for predicting the impact of risk on cost of building projects

Abstract

The use of construction project features (CPFs) to predict the impact of risk on costs of building projects is severely limited by the necessity to gather a homogeneous sample of projects. This limitation of the use of CPFs for risk prediction is the problem addressed by the study. The study aimed to develop artificial neural networks for predicting the occurrence, type and degree of impact of risk on costs of building projects by using selected CPFs. Data on 69 building projects was collected through the use of questionnaires from Quantity Surveyors in Abuja who were purposively sampled. The study found that costs of building projects are impacted by eight risks, which include variation, scope and design changes; error/omission in design/estimates, and unforeseen economic, site and social conditions. Project consultants are responsible for 69% of risks occurrence, while 52% of the cost impacts of risks result from the actions of clients. ANN1, an MLP artificial neural network with 2:31:1:1 structure was developed to predict variance between initial and final contract values by using five of the eight risks in two groups of client and consultant risks. A validation MSE of 0.0026 established ANN1’s superiority over a conventional MLR statistical model (Final cost variance = -4.834 + 1.056Consultant Risks + 1.058Client Risks) which had an MSE of 10.22. ANN2, an 8:19:7 MLP network was developed to predict risk effect on building costs by using 8 CPFs including gross floor area and costs of building elements. ANN2 used binarization to normalize data, with a resultant MSE of 0.2109, although lower MSE of 0.09 and higher specificity were obtained when risks were predicted one at a time. Optimum network settings for activation function, number of neurons and threshold were also derived for ANN2. The study concluded that using the derived network settings optimized network sensitivity, enabling ANN2 to correctly predict 9 out of 10 occurrences of risk, with a minimal false alarm rate of 2 out of 10. This is considered very satisfactory because clients are more interested in the occurrence of risk, which usually results in more money being needed to achieve ongoing projects. It was recommended that the developed networks ANN1 and ANN2 could be applied in the estimation of cost variance and risk effect on building costs, early in the construction phase when designs have been finalized but construction is yet to commence.