Abstract

This thesis presents a neural network-based cost estimating method, developed for the generation of conceptual cost estimates for low-rise prefabricated structural steel buildings. Detailed cost estimating is current practice for this type of buildings, since cost estimators are often challenged by a wide variety of different parameters. The developed method employs neural networks (NNs) for modeling individual project parameters associated with the direct cost of a project. It integrates NN cost models with cost adjustments, allowing for evaluation of different project alternatives, in a timely manner. The ability of NNs to capture real life experiences encountered on actual projects (i.e. actual costs), generalize and utilize that knowledge for estimating the cost of new projects makes it a very powerful tool to the application at hand. Data used in this study (75 building projects) were collected from a large manufacturer of prefabricated structural steel buildings in Canada (Canam Manac) over a 3-month period. The performance of developed cost models was tested against costs incurred by projects not used in training of those models, and costs predicted by regression. Results indicate that the proposed models, when used for projects with parameters within the range for which the models were trained, outperform regression. In addition, the proposed models can account for a number of parameters defining a project, and bearing considerable impact on the project cost. The proposed methodology can easily be adapted to provide decision-support for risk management and to assist in developing productivity models in a wide range of industries.