Evaluation of models to predict the construction material quantities of cylindrical storage structures

In the determination of whether or not a manufacturing plant should be constructed, it is necessary, during the early stages of a project, to make accurate estimates of how much the plant will cost. Most predictive models used in the preliminary estimation of plant costs have been focused directly on the costs, and therefore, have grouped together uncertainties related to the amount of construction materials used in construction, the construction processes employed, and the actual prices paid for these materials and the execution of these processes. In order to improve upon the accuracy of the prediction of plant costs at an early phase, the first uncertainty to be reduced is that associated with the construction material quantities (CMQs). A study including the as-built bill of quantities from 53 cylindrical storage structures (CSSs) of different types from 13 manufacturing plants around the world has been conducted to evaluate the performance of different parametric construction-material-quantity models using different techniques. Six models based on regression analysis, developed using different techniques, and one model based on Neural Networks, developed using Generalized Reduced Gradient (GRG) nonlinear optimization, were investigated. The models were constructed taking into consideration 13 potential independent variables (8 continuous and 5 categorical). The ability of the models to predict the quantity of concrete and reinforcement steel in CSSs was evaluated by comparison of the adjusted R2, the standard error of estimate, the mean absolute percentage error (MAPE) and plots of the cumulative probability vs. the percentage error. It was found that the best models for the prediction of the amount of concrete (m3) and the amount of reinforcement (tons) to be used in CSS construction were built using Neural Networks. The best regression models were the ones built using the backward elimination multiple regression technique. All the models created using different techniques met the expected accuracy ranges for Class 4 and Class 5 estimates proposed by the Association for the Advancement of Cost Engineering International.