N-BEATS-RNN: deep learning for time series forecasting

• Ano de defesa: 2021
• Autor(a) principal: Sbrana, Attilio
• Orientador(a): Naldi, Murilo Coelho
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus Sorocaba
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC-So
• Palavras-chave em Português: Previsão de séries temporais; Aprendizado de máquina; Aprendizado profundo; Time series forecasting; Deep learning; Machinel learning
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::MATEMATICA DA COMPUTACAO; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/13820

This work presents N-BEATS-RNN, an extended version of an ensemble of deep learning networks for time series forecasting, N-BEATS. We apply a state-of-the-art Neural Architecture Search, based on a fast and efficient weight-sharing search, to solve for an ideal Recurrent Neural Network architecture to be added to N-BEATS. We evaluated the proposed N-BEATS-RNN architecture in the widely-known M4 competition dataset, which contains 100,000 time series from a variety of sources. N-BEATS-RNN achieves comparable results to N-BEATS and the M4 competition winner while employing solely 108 models, as compared to the original 2,160 models employed by N-BEATS, when composing its final ensemble of forecasts. Thus, N-BEATS-RNN's biggest contribution is in its training time reduction, which is in the order of 9 times compared with the original ensembles in N-BEATS.