Spare parts intermittent demand forecastin
Data
2023-06-02
Autores
Imaniche, Carlos César Minoru 
Orientadores
Sato, Renato Cesar
Tipo
Dissertação de mestrado
Título da Revista
ISSN da Revista
Título de Volume
Resumo
As demandas intermitentes ocorrem com frequência no ambiente de peças de reposição de aeronaves e componentes, sendo elas um dos principais problemas enfrentados pelas organizações modernas gerando enormes desafios para um bom planejamento. Além disso, a responsabilidade de empregar os recursos financeiros disponibilizados para a manutenção aeronáutica da forma mais eficiente implicam na necessidade de desenvolvimento de ferramentas capazes de realizar o controle e planejamento cada vez mais precisos. Desta forma, este trabalho tem o objetivo de fornecer uma metodologia de previsão de demanda de peças de reposição robusta capaz de lidar com suas possíveis características intermitentes. Para isso, foram propostos quatro modelos que utilizam um pool métodos consagrados para realização de previsões. O primeiro modelo simula como tradicionalmente seria realiza a escolha de um método de previsão. O segundo realiza a classificação do melhor método através das feições das séries temporais e precisão da previsão dos métodos do pool, utilizando o sistema de ensemble Random Forest. O terceiro modelo também realiza a classificação do melhor método da mesma que o segundo modelo, porém utilizando o sistema de ensemble XGBoost. O quarto modelo realiza a regressão da previsão usando diretamente as feições das previsões no sistema de ensemble XGBoost. Os quatro modelos foram submetidos à três conjuntos de dados sintéticos simulados com diferentes percentuais de séries temporais intermitentes em sua composição. O quarto modelo se mostrou bem robusto, obtendo o menor RMSE médio nos três conjuntos de dados. Dentro do conhecimento deste autor, este trabalho é um dos primeiros trabalhos que utiliza um algoritmo de meta-aprendizagem específico para lidar com conjuntos de séries temporais com características intermitentes, sendo a principal contribuição fornecer uma nova ferramenta capaz de realizar previsões robustas para este tipo de conjunto de dados, em um baixo tempo de processamento computacional.
Intermittent demands often occur in the environment of aircraft spare parts and components, being one of the main problems faced by modern organizations creating enormous challenges for good planning. Furthermore, the responsibility to use the resources available for aircraft maintenance in the most efficient way implies the need to develop tools capable of performing increasingly precise control and planning. Consequently, this work aims to provide a robust spare parts demand forecasting methodology capable of dealing with its possible intermittent characteristics. To perform this, four models were proposed, using a pool of consecrated methods to forecasting. The first model simulates how it would traditionally perform the choice of a forecasting method. The second performs the classification of the best method through the features of the time series and the accuracy of the prediction of the pool methods, using the Random Forest ensemble system. The third model also performs the classification of the best method in the same way as second model, but using the XGBoost ensemble system. The fourth model performs the forecast by regression using directly the features of the forecasts in the XGBoost ensemble system. The four models were submitted to three synthetic datasets with different percentages of intermittent time series in their composition. The fourth model proved to be very robust, obtaining the lowest average RMSE in the three datasets. To the knowledge of this author, this work is one of the first that uses a specific meta-learning algorithm to deal with sets of time series with intermittent characteristics, the main contribution of this work is provide a new tool capable of making robust predictions for this type of dataset, in a low computational processing time.
Intermittent demands often occur in the environment of aircraft spare parts and components, being one of the main problems faced by modern organizations creating enormous challenges for good planning. Furthermore, the responsibility to use the resources available for aircraft maintenance in the most efficient way implies the need to develop tools capable of performing increasingly precise control and planning. Consequently, this work aims to provide a robust spare parts demand forecasting methodology capable of dealing with its possible intermittent characteristics. To perform this, four models were proposed, using a pool of consecrated methods to forecasting. The first model simulates how it would traditionally perform the choice of a forecasting method. The second performs the classification of the best method through the features of the time series and the accuracy of the prediction of the pool methods, using the Random Forest ensemble system. The third model also performs the classification of the best method in the same way as second model, but using the XGBoost ensemble system. The fourth model performs the forecast by regression using directly the features of the forecasts in the XGBoost ensemble system. The four models were submitted to three synthetic datasets with different percentages of intermittent time series in their composition. The fourth model proved to be very robust, obtaining the lowest average RMSE in the three datasets. To the knowledge of this author, this work is one of the first that uses a specific meta-learning algorithm to deal with sets of time series with intermittent characteristics, the main contribution of this work is provide a new tool capable of making robust predictions for this type of dataset, in a low computational processing time.
Descrição
Citação
IMANICHE, Carlos César Minoru. Spare Parts Intermittent Demand Forecasting. 2023. 104f. Dissertation of Master of Science – Instituto Tecnológico de Aeronáutica and Universidade Federal de São Paulo, São José dos Campos, 2023.