A carregar...
Publicação
Modelos de Machine Learning para a Deteção de Anomalias em Consumos Energéticos
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.26 MB | Adobe PDF |
Autores
Resumo(s)
O consumo de energia em edifícios institucionais é influenciado por múltiplos fatores e pode ser difícil de monitorizar de forma eficiente sem ferramentas adequadas de apoio à decisão. A previsão do consumo energético, quando realizada com elevada precisão, pode ser utilizada como referência de comportamento esperado, permitindo identificar desvios significativos que possam corresponder a anomalias operacionais. O presente trabalho, desenvolvido em continuidade com o trabalho de investigação de J. Sá, tem como objetivo o desenvolvimento de modelos de estimação do consumo diário de energia com um erro reduzido, de modo a poderem ser utilizados como suporte à deteção de anomalias no consumo energético. A metodologia seguida consiste na utilização de um modelo de previsão horária do consumo e na agregação dos valores previstos ao longo do dia, permitindo obter uma estimativa diária do consumo. Os resultados mostram que esta metodologia simples permite obter valores de erro extremamente baixos, deixando em aberto a possibilidade da sua utilização como suporte à deteção de anomalias no consumo. Foram utilizados os registos do Campus do Instituto Politécnico de Leiria, recolhidos entre 2015 e 2023, para avaliar o desempenho da metodologia proposta. Para tal, foram desenvolvidos e comparados seis modelos: MLP, LSTM, GRU, XGBoost, N-BEATS e N-HiTS. Os resultados demonstram que esta metodologia permite obter erros com MAPE médio entre 1,28% e 1,70%, evidenciando o bom desempenho preditivo dos modelos desenvolvidos.
Energy consumption in institutional buildings is influenced by multiple factors and can be challenging to monitor efficiently without adequate decision-support tools. Energy consumption forecasting, when performed with high accuracy, can serve as a reference for expected behavior, enabling the identification of significant deviations that may correspond to operational anomalies. The present study, developed as a continuation of the research conducted by J. Sá, aims to develop energy consumption estimation models with low error, so that they can be used to support anomaly detection in energy consumption. The methodology consists of using an hourly consumption forecasting model and aggregating the predicted values throughout the day, yielding a daily consumption estimate. The results show that this simple methodology achieves very low error values, indicating its potential use as support for anomaly detection in energy consumption. Hourly records from the Campus of the Polytechnic Institute of Leiria, collected between 2015 and 2023, were used to evaluate the performance of the proposed methodology. Six models were developed and compared: MLP, LSTM, GRU, XGBoost, N-BEATS and N-HiTS. The results demonstrate that this methodology achieves mean MAPE values between 1,28% and 1,70%, indicating the good predictive performance of the developed models.
Energy consumption in institutional buildings is influenced by multiple factors and can be challenging to monitor efficiently without adequate decision-support tools. Energy consumption forecasting, when performed with high accuracy, can serve as a reference for expected behavior, enabling the identification of significant deviations that may correspond to operational anomalies. The present study, developed as a continuation of the research conducted by J. Sá, aims to develop energy consumption estimation models with low error, so that they can be used to support anomaly detection in energy consumption. The methodology consists of using an hourly consumption forecasting model and aggregating the predicted values throughout the day, yielding a daily consumption estimate. The results show that this simple methodology achieves very low error values, indicating its potential use as support for anomaly detection in energy consumption. Hourly records from the Campus of the Polytechnic Institute of Leiria, collected between 2015 and 2023, were used to evaluate the performance of the proposed methodology. Six models were developed and compared: MLP, LSTM, GRU, XGBoost, N-BEATS and N-HiTS. The results demonstrate that this methodology achieves mean MAPE values between 1,28% and 1,70%, indicating the good predictive performance of the developed models.
Descrição
Palavras-chave
Consumo energético Estimação do consumo Séries temporais Machine learning Deteção de anomalias