| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 5.16 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Da procura pela melhoria de processos e serviço de excelência ao cliente, surgiu o enquadramento
para esta dissertação, que teve por base o trabalho desenvolvido no departamento de Após-venda
da Salvador Caetano África. O seu principal objetivo prendeu-se com a análise de valores de frete
pago em cada transporte de peças, com vista à criação de um sistema robusto que permitisse
auxiliar a previsão deste valor com base em parâmetros previamente conhecidos.
A dissertação inicia-se com uma revisão de literatura nas áreas de logística, modelação de
processos, melhoria contínua, sistemas de apoio à decisão, bem como conceitos associados ao
Machine Learning, séries temporais e modelos de previsão.
Deste modo, através da recolha e tratamento dos dados históricos existentes, foram aplicados uma
série de cálculos e métricas de análise de séries temporais, de forma a identificar o comportamento
das séries de dados em questão e, consequentemente, que tipo de metodologia poderia ser mais
indicada para prever valores. Posteriormente, aplicaram-se algoritmos específicos, testando-se
diferentes modelos de previsão, com o objetivo de se conseguir identificar qual deles apresentaria
melhor comportamento na previsão de valores futuros. Finalmente, analisou-se e elegeu-se aquele
que foi o modelo que teve o melhor ajuste, isto é, a capacidade de prever, com alguma exatidão,
valores futuros, gerando assim a menor percentagem de erro face a valores reais.
Os modelos foram testados para previsão de dados futuros, obtendo-se valores de erro entre os
17% e os 56% no caso do Alisamento Exponencial Simples (ETS) e, valores entre os 5% e 27%,
quando aplicados modelos híbridos de Deep Learning (BJ-DNN).
Através da utilização de modelos de previsão, bem como a recolha e seleção de parâmetros a
contribuir para o valor de frete, foi possível retirar conclusões sobre a validade dos mesmos, bem
como identificar passos futuros a serem implementados.
The search for process improvement and excellent customer service gave rise to the framework for this dissertation, which was based on work carried out in the After-Sales department of Salvador Caetano Africa. Its main objective was to analyze the freight paid for each shipment of parts, with a goal of creating a robust system to help predict this value based on previously known parameters. The dissertation begins with a literature review in the areas of logistics, process modelling, continuous improvement, decision support systems, as well as concepts associated with Machine Learning, time series and forecasting models. Thus, by collecting and processing existing historical data, a series of calculations and time series analysis metrics were applied in order to identify the behavior of the data series in hand and, consequently, what type of methodology might be the most suitable for predicting values. Specific algorithms were then applied, testing different forecasting models in order to identify which one would perform best in predicting future values. Finally, the model with the best fit was analyzed and chosen, i.e. the model with the ability to predict future values with some level of accuracy, thus generating the lowest percentage of error in relation to real values. The models were tested for predicting future data, obtaining error values between 17% and 56% in the case of Exponential Smoothing (ETS) models and values between 5% and 27% when hybrid Deep Learning (BJ-DNN) models were applied. Through the use of forecasting models, as well as the collection and selection of parameters to contribute to the freight value, it was possible to draw conclusions about their validity, as well as identify future steps to be implemented.
The search for process improvement and excellent customer service gave rise to the framework for this dissertation, which was based on work carried out in the After-Sales department of Salvador Caetano Africa. Its main objective was to analyze the freight paid for each shipment of parts, with a goal of creating a robust system to help predict this value based on previously known parameters. The dissertation begins with a literature review in the areas of logistics, process modelling, continuous improvement, decision support systems, as well as concepts associated with Machine Learning, time series and forecasting models. Thus, by collecting and processing existing historical data, a series of calculations and time series analysis metrics were applied in order to identify the behavior of the data series in hand and, consequently, what type of methodology might be the most suitable for predicting values. Specific algorithms were then applied, testing different forecasting models in order to identify which one would perform best in predicting future values. Finally, the model with the best fit was analyzed and chosen, i.e. the model with the ability to predict future values with some level of accuracy, thus generating the lowest percentage of error in relation to real values. The models were tested for predicting future data, obtaining error values between 17% and 56% in the case of Exponential Smoothing (ETS) models and values between 5% and 27% when hybrid Deep Learning (BJ-DNN) models were applied. Through the use of forecasting models, as well as the collection and selection of parameters to contribute to the freight value, it was possible to draw conclusions about their validity, as well as identify future steps to be implemented.
Description
Keywords
Logística Modelação de Processos Previsão Alisamento Exponencial Deep Learning