A carregar...
Publicação
Análise de séries temporais de absentismo no setor ferroviário
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 6.65 MB | Adobe PDF |
Orientador(es)
Resumo(s)
O absentismo constitui um problema para os operadores ferroviários, pois pode ter
repercussões, como, por exemplo, o cancelamento de viagens. Pela natureza do serviço
que prestam, os operadores ferroviários efetuam planeamentos operacionais. Porém, como
alguns tipos de ausências dos tripulantes, nomeadamente aquelas motivadas por razões
de doença, não se conseguem prever, os referidos operadores procuram outras ferramentas
de apoio à decisão que possam ajudar a prevenir os impactos provocados pelo absentismo.
Esta Dissertação tem como finalidade desenvolverummodelo de previsão de ausências
não programáveis que tenha capacidade de estimar o número de tripulantes ausentes por
motivo de doença e/ou apoio à família, com base no histórico de atividade planeada e
realizada dos tripulantes, para que possa ser englobado no planeamento.
Os dados foram disponibilizados por parte de um operador ferroviário da Europa
do Norte, um cliente da empresa SISCOG. Neste caso, analisaram-se dados diários sobre
a percentagem de maquinistas e de revisores ausentes de uma base operacional
grande. Realizou-se um pré-processamento e uma análise exploratória, para perceber as
características de cada série.
No final, foram comparados os resultados entre o modelo regressão dinâmica harmónica
(RDH) e os algoritmos Prophet e NeuralProphet, com o método da janela fixa e
método da janela móvel. Quanto à janela fixa, o NeuralProphet destacou-se em termos de
precisão para a série dos revisores, com um MAPE de 12.86%. Para a série dos maquinistas,
não se conseguiu concluir qual o método mais adequado, uma vez que as métricas eram
muito semelhantes. No entanto, sugere-se que o Prophet, com um MAPE de 20.70%,
seria a escolha mais indicada para a série devido ao seu custo computacional. Porém, os
resultados da janela móvel revelaram que o modelo de RDH obteve um bom desempenho
ao longo dos diferentes horizontes de previsão, em comparação com o NeuralProphet.
Quanto aos dados dos maquinistas, os modelos obtiveram um MAPE pequeno no último
horizonte de previsão, em comparação com os restantes.
Absenteeism is a problem for railway operators because it can have repercussions such as cancellation of trips. Due to the nature of the service delivered, railway operators carry out operational planning. However, as some types of absences of crew members, particularly those caused by illness, cannot be predicted, they seek other decision support tools that can help prevent the impacts caused by absenteeism. This dissertation aims to develop a prediction model for unscheduled absences that can estimate the number or percentage of crew members unexpectedly absent due to illness and/or family support, for a given time period, based on the planned and performed activity history of the crew members, so that it can be incorporated into the planning. The data was made available from a railway operator in Northern Europe, a customer of the SISCOG company. In this case, we analyzed data on the percentage of absent train drivers and guards from selected operational bases. Data was pre-processed, and then an exploratory analysis was carried out, to understand the characteristics of each series. Finally, the results produced by the dynamic harmonic regression (DHR) model and Prophet and NeuralProphet algorithms were compared, by using the fixed window and sliding window methods. Regarding the fixed window method, the NeuralProphet stood out for its precision for the series related to the guards, with a MAPE of 12.86%. For the series related with train drivers we could not conclude which method was the most appropriate, as the metrics were very similar. However, we decided that Prophet, with a MAPE of 20.70%, would be the most appropriate choice for this series, because of its computational cost. Additionally, the sliding window results revealed the DHR model as having a good performance along the diferent prediction horizons, when compared with NeuralProphet. regarding the train drivers data, the models obtained a small MAPE value in the last prediction horizon, compared to the less distant prediction horizons.
Absenteeism is a problem for railway operators because it can have repercussions such as cancellation of trips. Due to the nature of the service delivered, railway operators carry out operational planning. However, as some types of absences of crew members, particularly those caused by illness, cannot be predicted, they seek other decision support tools that can help prevent the impacts caused by absenteeism. This dissertation aims to develop a prediction model for unscheduled absences that can estimate the number or percentage of crew members unexpectedly absent due to illness and/or family support, for a given time period, based on the planned and performed activity history of the crew members, so that it can be incorporated into the planning. The data was made available from a railway operator in Northern Europe, a customer of the SISCOG company. In this case, we analyzed data on the percentage of absent train drivers and guards from selected operational bases. Data was pre-processed, and then an exploratory analysis was carried out, to understand the characteristics of each series. Finally, the results produced by the dynamic harmonic regression (DHR) model and Prophet and NeuralProphet algorithms were compared, by using the fixed window and sliding window methods. Regarding the fixed window method, the NeuralProphet stood out for its precision for the series related to the guards, with a MAPE of 12.86%. For the series related with train drivers we could not conclude which method was the most appropriate, as the metrics were very similar. However, we decided that Prophet, with a MAPE of 20.70%, would be the most appropriate choice for this series, because of its computational cost. Additionally, the sliding window results revealed the DHR model as having a good performance along the diferent prediction horizons, when compared with NeuralProphet. regarding the train drivers data, the models obtained a small MAPE value in the last prediction horizon, compared to the less distant prediction horizons.
Descrição
Palavras-chave
Absentismo Operador ferroviário Planeamento operacional Tripulantes Métodos de Previsão Séries temporais