Aprendizado de máquina com treinamento continuado aplicado à previsão de demanda de curto prazo: o caso do restaurante universitário da Universidade Federal de Uberlândia

• Ano de defesa: 2019
• Autor(a) principal: Silva, Leonildo Costa e
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Engenharia Elétrica
• Palavras-chave em Português: Planejamento; Planning; Aprendizado de máquina; Machine learning; Treinamento continuado; Continuous learning; Teoria da ressonância adaptativa; Adaptive resonance theory; Previsão de demanda; Demand forecasting; Redes neurais artificiais; Artificial neural network; Engenharia elétrica; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufu.br/handle/123456789/25386; http://dx.doi.org/10.14393/ufu.di.2019.2001

Planning in situations of uncertainty requires the use of information that assists managers in decision making. The analysis of time series with the aim of creating predictions about future events is an important ally in this process. In this context the demand forecast analysis can be defined as a projection of the future on the demand for goods or services offered by an organization. With this information, managers can guide their actions in function of the operational and strategic objectives. Therefore, generating reliable and accurate forecasts is an indispensable activity for any institution focused on the supply of goods or services. This work investigates the application of methods based on machine learning for the generation of short term quantitative forecasts. Given the volatility characteristic of consumer relations influenced by several factors, the relationship between the explanatory variables and the forecast demand may change over time. Thus, forecasting models may show degradation of their accuracy if they are not constantly updated. Thus, this research compares the performance of four machine learning algorithms with continuous training capacity. The dataset used in the experiment consists of the daily demand for meals at the university restaurant of the Federal University of Uberlândia. To train the models, data were collected between January 2015 and April 2018, and the last six months were reserved for the test procedure. The methodology adopted is based on the CRISP-DM process with the use of appropriate practices for machine learning projects. The results indicated that the methodology employed was able to produce models with good quality of fit (uncorrelated residues and residual mean close to zero) and an adequate generalization capacity, measured by RMSE and MAE metrics.