Estimation of construction waste generation using machine learning

Abstract

The estimation of construction and demolition waste is a difficult task, because it depends on technical, cultural and geometric variables of the buildings. Machine learning has been increasingly used in the construction industry, because it combines the calculation of large amounts of data with the difficulty of describing or understanding construction techniques related to human behaviour. This study evaluated the performance of artificial neural networks on predicting the amount of waste generated in construction works. Through an exploratory research and analysis of 330 works, the performance of neural networks with two, five or ten neurons in the hidden layer, using gross floor area and working duration as input data has been studied. Three training algorithms were tested. Up to five training cycles were simulated. The results showed that the most appropriate training algorithm for the case is Bayesian regularisation and, using two neurons in the hidden layer and two training cycles, excellent prediction results can be achieved, with R2 of 0.96. The best configuration proposed for the neural network was able to accurately predict 43% of cases. The study showed that results obtained from machine learning worked better than those obtained from linear multiple regression models, usual in the literature.