A carregar...
Publicação
Previsão de Carga Agregada no Horizonte do Planeamento Operacional
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 5.67 MB | Adobe PDF |
Orientador(es)
Resumo(s)
O estudo deste trabalho provém da intenção de adquirir valores de previsão
de carga, com uma melhor performance relativamente aos casos atuais da litera-
tura, utilizando técnicas de previsão de carga agregada para um horizonte de
planeamento operacional (24h-72h). Este estudo, tem como base um conjunto de
dados relativos ao consumo energético português ao longo de 8 anos (2009 a
2017).
Para atingir este objetivo, foram utilizados dois algoritmos: um método de
agrupamento de dados, os mapas auto-organizados (SOM) e um método de re-
gressão, as redes neuronais artificiais multicamadas (MLP). Através dos SOM’s,
foi possível identificar similaridades entre dias, meses e certas alturas do ano que,
posteriormente, permitem a formação de grupos com características semelhan-
tes. Através deste, foram formados 24 grupos que, posteriormente, para cada
grupo, foi treinada uma rede MLP, para permitir a previsão da carga para dias
pertencentes a cada grupo de dados. A junção destes dois algoritmos faz destes
um modelo híbrido.
Este modelo híbrido (SOM + MLP), permitiu-nos obter um MAPE anual de
3.17%. Tendo em conta que este trabalho não utilizou valores relativos a dias an-
teriores (esta informação pode não estar disponível no dia em que se pretende
realizar a previsão), este valor de MAPE verifica-se abaixo dos valores da litera-
tura atual, o que permite concluir que a escolha deste modelo é adequada ao pro-
blema. Foi ainda possível identificar os dias que mais contribuíam para o au-
mento do erro da previsão, por serem de difícil alocação num dos grupos estabe-
lecidos. Para estes dias, é recomendado utilizar técnicas de previsão dedicadas
aos mesmos.
The study of this work comes from the intention of acquiring load forecast- ing values, with a better performance in relation to current cases in the literature, using aggregated load forecasting techniques for an operational planning hori- zon (24h-72h). This study is based on a dataset regarding the Portuguese energy consumption over 8 years (2009 to 2017). To achieve this goal, two algorithms were used: a clustering method, self- organizing maps (SOM) and a regression method, multilayer perceptron (MLP). Through the SOM's, it was possible to identify similarities between days, months and certain times of the year that, subsequently, allow the formation of groups with similar characteristics. Through this algorithm, 24 groups were formed and, subsequently, for each group, an MLP network was trained in order to obtain the load forecast for the days belonging to each group of data. The combination of these two algorithms makes them a hybrid model. This hybrid model (SOM + MLP) allowed us to obtain an annual MAPE of 3.17%. Considering that this work did not use values from previous days (this information may not be available on the day we intend to perform the forecast), this MAPE value is below the values in current literature, which allows us to conclude that the choice of this model is adequate to the problem. It was also possible to identify the days that most contribute to the increase of the forecast error, due to the difficulty to allocate to one of the established groups. For these days it is recommended to use dedicated forecasting techniques dedicated to them.
The study of this work comes from the intention of acquiring load forecast- ing values, with a better performance in relation to current cases in the literature, using aggregated load forecasting techniques for an operational planning hori- zon (24h-72h). This study is based on a dataset regarding the Portuguese energy consumption over 8 years (2009 to 2017). To achieve this goal, two algorithms were used: a clustering method, self- organizing maps (SOM) and a regression method, multilayer perceptron (MLP). Through the SOM's, it was possible to identify similarities between days, months and certain times of the year that, subsequently, allow the formation of groups with similar characteristics. Through this algorithm, 24 groups were formed and, subsequently, for each group, an MLP network was trained in order to obtain the load forecast for the days belonging to each group of data. The combination of these two algorithms makes them a hybrid model. This hybrid model (SOM + MLP) allowed us to obtain an annual MAPE of 3.17%. Considering that this work did not use values from previous days (this information may not be available on the day we intend to perform the forecast), this MAPE value is below the values in current literature, which allows us to conclude that the choice of this model is adequate to the problem. It was also possible to identify the days that most contribute to the increase of the forecast error, due to the difficulty to allocate to one of the established groups. For these days it is recommended to use dedicated forecasting techniques dedicated to them.
Descrição
Palavras-chave
mapas auto-organizados redes neuronais artificiais multi- camadas séries temporais previsão de carga modelos híbridos de previsão