Abstract:
In electrical power systems, one of the main challenges faced by system operators (ISO) is
ensuring a reliable balance between energy supply and demand. For this, a proper planning of
the system operation is necessary. Short-Term Load Forecasting (PCCP) is essential in this
process, as it assists in planning daily operations, including matching demand and supply,
setting future electricity prices and defining generation reserves. As smart grid technologies
and intermittent renewable energy sources have increased significantly in electricity markets,
the task of forecasting the load has increased in complexity and poses a challenge for ISOs.
Nonlinear models based on machine learning techniques have become quite popular in recent
years, as has the use of hybrid models focused on specific problems. This work focuses on the
development of artificial neural network models to solve the PCCP problem, among them the
MultiLayer Perceptron (MLP) networks with different numbers of layers and the Long-Short
Term Memory (LSTM) and Gated Recurrent Unit (GRU) recurrent networks in its
unidirectional and bidirectional versions. In addition to the data provided by ISO, predicted
temperature information obtained from the Global Ensemble Forecast System (GEFS) model
was used to try to generate more accurate load forecasts. In general terms, the results show that
the recurrent models produce greater accuracy and more reliable results than the other models,
including the models used by the Brazilian ISO. This is emphasized with the application of the
Diebold-Mariano in a paired comparison test between models.