An Expanded Study of the Application of Deep Learning Models in Energy Consumption Prediction

The time series of electrical loads are complex, influenced by multiple variables (endogenous and exogenous), display non-linear behavior and have multiple seasonality with daily, weekly and annual cycles. This paper addresses the main aspects of demand forecast modeling from time series and applies machine learning techniques for this type of problem. The results indicate that through an amplified model including the selection of variables, seasonality representation technique selection, appropriate choice of model for database (deep or shallow) and its calibration, it's possible to archive better results with an acceptable computational cost. In the conclusion, suggestions for the continuity of the study are presented.