Metodologia baseada em aprendizado profundo para agrupamento de séries temporais univariadas

• Ano de defesa: 2021
• Autor(a) principal: ARAÚJO, Alexandre de Carvalho
• Orientador(a): PAIVA, Anselmo Cardoso de
• Banca de defesa: PAIVA, Anselmo Cardoso de , SOUSA, João Dallyson Sousa de Almeida de , GOMES JÚNIOR, Daniel Lima , SERRA, Ginalber Luiz de Oliveira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: Agrupamento de séries temporais; Aprendizado profundo; Estimação de número de grupos
• Palavras-chave em Inglês: Time series clustering; Deep Clustering; Cluster number estimation; Electrical energy consumption
• Área do conhecimento CNPq: Ciência da Computação
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/3354

Electrical energy companies have a need to plan energy distribution to fulfill the needs of it’s costumers. They also need to analyze energy consumption to detect irregularities. To perform these tasks, consumption patterns discovery is an essential tool. Discovery of these patterns is a clustering task. There is a great number of approachs to cluster data, but a relatively new research subject for this task is the usage of deep learning. Great effort in this area is allocated to static data. However, applications that use time series data have been growing in the last years and not as much research effort is being made in this area. In this work, we propose a new methodology based in the Deep Embedded Clustering architecture, a popular deep clustering algorithm, to cluster electric consumption time series data while also approaching one of the flaws present on DEC, the need to inform number of clusters a priori, which isn’t generally known in real problems. To evaluate our methodology, we conducted tests regarding cluster number estimation in 66 time series datasets with known cluster numbers and to evaluate the consumption pattern discovery, we use 2 real electrical consumption datasets provided by a energy consumption company. Our results show superior performance when compared to approaches found in the literature for both cluster number estimation and clustering, indicating that the proposed methodology may be a effective tool for electrical energy companies.