Loading...
Publication
Comparação de Modelos de Previsão de Séries Temporais - Aplicação a Caudais de Água Distribuída em Redes de Abastecimento
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 13.89 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Este trabalho visa efetuar uma comparação entre modelos de previsão de séries temporais, nomeadamente, Holt-Winters, ARIMA, Long Short-Term Memory (LSTM) e Prophet, quando aplicados a dados reais de caudais de água distribuídos em quatro setores de abastecimento de aglomerados urbanos, situados em zonas rurais, onde existia uma medição de caudal e respetivo registo com uma frequência máxima de uma hora. Os setores em causa foram o de Janeiro de Cima e Aldeia de Joanes, no concelho do Fundão, Degolados, no concelho de Campo Maior e Alcáçova, no concelho de Elvas, todos situados em Portugal. A comparação foi feita de forma determinística utilizando as métricas de avaliação Erro Médio Absoluto, Raiz do Erro Quadrático Médio e Coeficiente de Determinação. As métricas foram calculadas para previsões de curto prazo, 10 dias, e de longo prazo, 3 meses. Os resultados mostraram que o modelo ARIMA era mais eficiente para previsões de curto prazo, sendo o modelo LSTM mais eficiente nas previsões de longo prazo. Verificou-se que em zonas com oscilações de consumo significativas que perturbavam os padrões de uma forma aleatória, nenhum modelo conseguiu efetuar previsões minimamente aceitáveis. Por outro lado, verificou-se a potencialidade dos modelos LSTM e Prophet em efetuarem previsões tendo em conta a sazonalidade anual, característica determinante para deteção de anomalias.
This work aims to perform a comparison among time series prediction models, namely Holt-Winters, ARIMA, Long Short-Term Memory (LSTM), and Prophet, when applied to real water flow data distributed in four supply sectors of urban clusters located in rural areas. These sectors include Janeiro de Cima and Aldeia de Joanes in the Fundão municipality, Degolados in Campo Maior municipality, and Alcáçova in Elvas municipality, all situated in Portugal. The water flow was measured and recorded with a maximum frequency of one hour. The comparison was conducted deterministically using evaluation metrics such as Mean Absolute Error, Root Mean Square Error, and Coefficient of Determination. The metrics were calculated for short-term forecasts (10 days) and long-term forecasts (3 months). The results showed that ARIMA model was more efficient for short-termpredictions, while the LSTM model was more efficient for long-term predictions. It was observed that in areas with significant consumption fluctuations that disturbed patterns randomly, no model could make minimally acceptable predictions. On the other hand, the potential of LSTM and Prophet models to make predictions considering annual seasonality, a decisive characteristic for anomaly detection, was confirmed.
This work aims to perform a comparison among time series prediction models, namely Holt-Winters, ARIMA, Long Short-Term Memory (LSTM), and Prophet, when applied to real water flow data distributed in four supply sectors of urban clusters located in rural areas. These sectors include Janeiro de Cima and Aldeia de Joanes in the Fundão municipality, Degolados in Campo Maior municipality, and Alcáçova in Elvas municipality, all situated in Portugal. The water flow was measured and recorded with a maximum frequency of one hour. The comparison was conducted deterministically using evaluation metrics such as Mean Absolute Error, Root Mean Square Error, and Coefficient of Determination. The metrics were calculated for short-term forecasts (10 days) and long-term forecasts (3 months). The results showed that ARIMA model was more efficient for short-termpredictions, while the LSTM model was more efficient for long-term predictions. It was observed that in areas with significant consumption fluctuations that disturbed patterns randomly, no model could make minimally acceptable predictions. On the other hand, the potential of LSTM and Prophet models to make predictions considering annual seasonality, a decisive characteristic for anomaly detection, was confirmed.
Description
Keywords
séries temporais modelos de previsão caudais distribuídos Holt- Winters ARIMA LSTM Prophet time series prediction models water flow distribution Holt-Winters