Forecasting aggregated hourly electricity demand in Southeast and Midwest Brazil

This study investigates aggregated hourly electricity demand from the Southeast and Midwest regions of Brazil, based on publicly available data from the National Electric System Operator. These two regions together account for approximately 56% of the country’s total electricity consumption, making the aggregated series highly representative of national demand patterns. The data exhibit clear multiple seasonalities, including daily, weekly, and annual cycles. The predictive performance of a range of forecasting models was evaluated using a rolling window framework over a dataset comprising eleven years of observations (2010–2020). Models were selected and fitted using data from 2010 to 2018, while forecast accuracy was assessed using more than 17,000 hourlly observations from 2019 to 2020. The evaluation considered forecasting horizons ranging from 1 to 168 h. The results demonstrate that models incorporating multiple seasonal patterns yield superior accuracy for forecasting horizons up to 24 h. For longer horizons, models focusing exclusively on the weekly seasonal cycle tend to perform better. Specifically, the SARIMA model capturing daily and weekly seasonalities achieves the highest accuracy in short-term forecasts, whereas the LSTM model configured with weekly seasonality provides better performance for extended horizons.