| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 2.28 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
A análise de séries temporais estão sempre presentes nos mais diferentes domínios, pela
grande facilidade de representar os acontecimentos em relação ao tempo. Alguns dos fatores
que podem ser investigados nas séries temporais são a descrição do comportamento da
série e/ou a realização de previsões dos valores futuros da série. O principal objetivo desta
dissertação foi encontrar um modelo preditivo que conseguisse apresentar valores futuros
satisfatórias para as tomadas de decisões empresariais. A metodologia utilizada consistiu
de três etapas principais que foram: a definição do problema, seguido da preparação dos
dados com uma investigação sobre estes dados para então criar um modelo preditivo e
finalmente apresentar os resultados. Identificou-se, na investigação sobre os dados, que
grande parte dos produtos não estavam mais sendo produzidos, sendo assim, foi utilizado
do k-means para realizar o agrupamento dos produtos com demanda regular e assim prosseguir
com a criação do modelo preditivos, em que, onde foram testados os algoritmos
de média móvel integrada autoregressiva sazonal (SARIMA), perceptron multicamadas
(MLP) e floresta aleatória (RF). Os melhores resultados foram apresentados pelo algoritmo
de Floresta Aleatória, com um coeficiente de determinação de aproximadamente
88%, com um erro absoluto médio de aproximadamente 189 unidades de produtos.
The analysis of time series are always present in the most different domains, due to the great ease of representing events in relation to time. Some of the factors that can be investigated in the time series are the description of the behavior of the series and/or the realization of forecasts of future values of the series. The main objective of this dissertation was to find a predictive model that could present satisfactory future values for business decision making. The methodology used consisted of three main steps that were: the definition of the problem, followed by the preparation of the data with an investigation on these data to then create a predictive model and finally present the results. It was identified in the research on the data that most of the products were no longer being produced, so that k-means was used to group products with regular demand and thus to proceed with the creation of the predictive models, where the algorithms of integrated autoregressive seasonal moving average (SARIMA), multilayer perceptron (MLP) and random forest (RF) were tested. The best results were presented by the Random Forest algorithm, with a determination coefficient of approximately 88%, with an mean absolute eror of approximately 189 product units.
The analysis of time series are always present in the most different domains, due to the great ease of representing events in relation to time. Some of the factors that can be investigated in the time series are the description of the behavior of the series and/or the realization of forecasts of future values of the series. The main objective of this dissertation was to find a predictive model that could present satisfactory future values for business decision making. The methodology used consisted of three main steps that were: the definition of the problem, followed by the preparation of the data with an investigation on these data to then create a predictive model and finally present the results. It was identified in the research on the data that most of the products were no longer being produced, so that k-means was used to group products with regular demand and thus to proceed with the creation of the predictive models, where the algorithms of integrated autoregressive seasonal moving average (SARIMA), multilayer perceptron (MLP) and random forest (RF) were tested. The best results were presented by the Random Forest algorithm, with a determination coefficient of approximately 88%, with an mean absolute eror of approximately 189 product units.
Description
Mestrado de dupla diplomação com a UTFPR - Universidade Tecnológica Federal do Paraná
Keywords
Séries temporais Aprendizado de máquina Previsão