Título: Comparative Analysis of Deseasonalization and Detrending Methods in Energy Consumption Forecasting
Autores: Taina de Souza Coimbra, Levy Boccato, Hugo Valadares Siqueira, Romis Attux
Resumo: Energy consumption forecasting is a valuable tool for management decision-making that can lower expenses and enhance efficiency in a power system. Several time series prediction models can address this issue, ranging from statistical models to complex neural networks. With the purpose of increasing prediction accuracy, we tackle the problem of time series stationarity by considering an Autoregressive (AR) and Multilayer Perceptron (MLP) model. This paper presents a comparative analysis of both models regarding different deseasonalization and detrending methods, which converts the time series to stationary. This is a necessary condition for linear statistical models such as Autoregressive (AR). Our focus is to investigate whether it also improves neural network performance and which stationarity method produces the best result. For this study, we address an energy consumption series from the southeast region of Brazil. The computational results reveal that removing the trend by differencing and removing seasonality by normalization leads to the lowest errors.
Palavras-chave: Energy consumption prediction, Auto-regressive Model, Neural Networks, Time Series, Stationarity, Seasonality, Trend.
Páginas: 7
Código DOI: 10.21528/CBIC2023-105
Artigo em pdf: CBIC_2023_paper105.pdf
Arquivo BibTeX: CBIC_2023_105.bib