Predição de demanda de energia elétrica para o dia seguinte aplicando redes neurais convolutivas e redes recorrentes

Abstract

Predicting energy demand in adverse scenarios, such as the COVID-19 pandemic, is critical to ensure the electricity supply and operation of basic services in a metropolitan region. Another important factor is the self-sustainability of the energy system, compared to historical levels of CO2 emissions caused by high energy consumption. In the literature, it is possible to observe a trend in the use of artificial intelligence to predict energy demand and mitigate these effects. However, few works involving anomalous scenarios have been developed. In this dissertation, a deep learning model is proposed to predict the demand for the next day using the “IEEE DataPort Competition Day-Ahead Electricity Demand Forecasting: Post-COVID Paradigm” database. Three deep neural network architectures are proposed, being a convolutional network, a recurrent network, and a hybrid (convolutional-recurrent) network. The best model presented is characterized by extracting spatio-temporal features from the input data by means of the hybrid network. A preliminary analysis of the input data is performed, with the exclusion of anomalous variables and the application of a sliding window, with a 24-hour interval, which defines the number of hours needed to predict the demand of the next hour. The normalization of the input data was also implemented, with a weighting with a factor of 10 for the demand variable. The performance of the proposed models was compared with the models developed in the competition, by means of a benchmark analysis. The hybrid architecture proposed in this work presented an average absolute error of 2361.84 kW (78.22% lower than the model with the best performance in the competition) and a similarity between the real demand curve and the predicted demand curve, also superior to the referred models. Such facts demonstrate the efficiency of the application of deep networks, compared to the classical methods applied by other authors.