Estimativa da quantidade de resíduos da construção civil utilizando redes neurais artificiais

• Ano de defesa: 2023
• Autor(a) principal: Recalcatti, Sandileia
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Engenharia Civil; UTFPR
• Palavras-chave em Português: Gestão integrada de resíduos sólidos; Aprendizado de máquina; Inteligência artificial; Construção civil - Estimativas - Curitiba (PR); Redes neurais (Computação); Resíduos como material de construção; Integrated solid waste mangement; Machine learning; Artificial intelligence; Building - Estimates - Curitiba (Brazil); Neural networks (Computer science); Waste products as building materials; CNPQ::ENGENHARIAS::ENGENHARIA CIVIL::CONSTRUCAO CIVIL; Engenharia Civil
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/31514

Waste generation is one of the negative impacts caused by civil construction, and faced by the public authorities and construction companies. Estimating the amount of civil construction waste (CCW) is fundamental in the management process, however, this task is not simple, because there are several materials and construction processes that can be used in the same enterprise. Builders are in search of quick solutions in civil construction, but many of the available quantification methods do not meet their needs adequately. Artificial neural networks (ANNs) can meet this need, due to their ability to learn and solve non-linear systems. In this sense, the general objective of this research is to investigate the use of ANN’s to estimate the generation of CCW in civil construction. The research adopted the simulation method, using samples A and B with 5,000 and 10,000 dummy data, respectively, both with total construction areas between 75 m² and 125,050 m², and sample R with data from 360 construction sites, with areas between 906 m² and 138,824 m². The dummy data was created based on generation rates available in the literature, and the real data was obtained through contact with construction companies located in Curitiba/PR, and covers buildings built between 2006 and 2021. To perform the simulations, the software MATLAB® version R2022a was used. Different configurations of neural networks were trained with samples A and B, and it was possible to verify that the best predictive result was in the training of sample B, with the feed-forward neural network with two input variables (waste classification and total built-up area), ten neurons in the hidden layer, one output variable (amount of waste) and three training cycles with the Bayesian Regularization algorithm, presenting R² values equal to 1.0, MSE equal to 42.87 kg and MAPE equal to 0.00013%. In the validation of the model with the R sample, the result of R² equal to 0.83 indicated a good performance of the neural network in explaining the variation of the output data based on the input data. In addition, the proposed model presented an MSE of 4,337.69 m³ and the MAPE result pointed out that the presented neural network model is able to accurately estimate more than 60% of the cases, as well as an optimal estimation presented by the ANN model when compared to other models in the literature. For this research, the neural network model that presented the best prediction results was the feed-forward neural network with ten neurons in the hidden layer and three training cycles with the Bayesian Regularization algorithm. This research brings an important contribution to the civil construction sector, collaborating with the quantification of CCW in an agile manner, besides contributing to the management of waste inside and outside the construction site, as well as contributing to the awareness of professionals in the area and is useful to base actions that minimize waste generation.