Use of virtual load curves for the training of neural networks for residential electricity consumption forecasting applications

Smart grids are becoming increasingly closer to consumers, especially residential consumers, bringing with them a wide range of possibilities. The level of information obtained on a smart grid will be much higher when compared to a traditional network and at this point, more informed consumers tend to consume more efficiently, bringing benefits to themselves and to the system. An interesting fact for control within a residence is forecasting consumption, allowing the consumer to know in advance how much to consume up to a certain period. Artificial neural networks are one of several methods used to forecast time series, however, require a high volume of historical data for the training of the network, given that these may not be accessible or even exist. At this point, the objective of this work is to evaluate the use of load curves obtained through computational tools for the pre-training of artificial neural networks used in the consumption forecast. A tool is used to create random load curves according to the region and socioeconomic characteristics. The load curves are transformed into cumulative consumption curves and used as training vectors of the artificial neural network. The results of the tests were very promising, they showed that the pretraining with the virtual data makes possible the forecast of the time series even in the absence of real data for the training, showing that the methodology developed has great potential of application in works related to the forecast consumption.