The project consists of the development of a Short-term Load forecasting tool to be used to prepare the distribution of Energy for the Electric System of Guatemala. The short term in this case is established to be the next day and 10 days ahead. The input data consists of a historical database of electricity Load for each hour, along with predictors such as temperature, influence of natural light and the operation of important loads such as the Electric Arc Furnace (EAF).
In this Model, the load is pre-processed to create predictors using a wide range of data, ranging from the type of day, month, year, hour, lagged load, moving averages and a special classification by type of holiday (December 24 and 25, Easter, November 1, etc.).
The User has the choice between two different Machine Learning and Deep Learning models implemented in MATLAB, to generate load forecast with great precision considering past data and future estimation of temperature, EAF operation and position of the sun in degrees relative to the earth plane. In this way, each time the tool is executed a new model is created, which is trained on the most recent data of the database, making the tool to never expire or deteriorate, but on the contrary, each time the model is used, it will be more robust and accurate.
The use of this tool represents a reduction of the absolute mean percentage error (MAPE) of 1.9% for estimation of the following day and 2.1% for the weekly projection, reducing by approximately 1.7% the error obtained using traditional techniques.