Intelligent assistant for electricity management

• Autor(a) principal: Oliveira, Pedro Miguel Rocha
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/38929
• topic: Anomaly detection; Deep learning; Machine learning; Meta-learning; Time series forecasting

The steady increase in energy consumption, caused by population growth worldwide, made energy forecasting to be a great tool for governmental and business decision making, such as electricity production or fuel trade. On one hand, accurately following the energy trend to predict how much electricity is expected to be used prevents wasteful spending. On the other hand, being able to detect early signs of equipment malfunction will also prevent additional expenditure. Therefore, creating an agent that’s both capable of predicting the energy to be used soon as well as detect anomalies would be highly helpful. This work describes the development of an intelligent agent to forecast an electricity data time series and use these predictions to detect anomalies in current time. The development focused on using various architectures featuring popular models for time series forecasting, such as LSTM and GRU, attention mechanisms and even using meta-learning models to create the best performing model. With these architectures, it was possible to create a model highly capable of forecasting the complex dataset of over 50 features, getting an average of 0.04 in normalized RMSE in the most relevant variables, and detect anomalies with high precision and reasonable recall, getting a value of 97.3% of precision for the same variables.