Short-term Load Demand Forecasting with Machine Learning Methods: A Brazilian Case Study

Accurate estimation and characterization of the load demand curve are critical tasks during the planning stage to ensure a secure and efficient power system operation. The aim is to coordinate the energy supply and the electricity market by predicting demand behavior, which must align with the scheduled generation levels at power plants. This paper identifies the most effective methods for short-term load demand forecasting across Brazil’s four regions, each with unique consumption patterns due to their size, using a one-week-ahead direct multi-step approach. The study evaluates the performance of the most representative machine learning methods from different categories, including linear regression (LR), multi-layer perceptron (MLP), gated recurrent unit (GRU), random forest (RF), and support vector regression (SVR). The dataset used includes the measured historical consumption data from Brazil, spanning 2016 to 2023, provided by the national electricity operator (ONS), and considers the four regions of the country. Forecast accuracy is assessed with the mean absolute percentage error (MAPE) and root mean squared error (RMSE). Results show that MLP and GRU outperform other methods, offering consistent performance across all regions and more reliable error estimates. These findings lay the groundwork for developing more robust tools, possibly through deeper architectures or model combinations.