Previsão de demanda de energia elétrica residencial utilizando aprendizado profundo

Abstract

Predicting a future electrical energy demand of a residence gets several applications on many scales. Starting with the energy supply quality indicator, going through the best management of dispatch on interconnect energy system, and getting to the bottom of expansion investments. Infer the demand takes a significant role in developing the energy segment. Therefore, certain methods help to infer consumption on various time horizons. Statistical models, support vector machines, and decision trees are some of the ways to solve the task, but deep neural networks come with a great focus on decreasing the error on the prediction, in which multiple features are presented with the consumption reading. Due to the recurrent neural network permissibility of the time-series prediction problem, it becomes interesting to compare these networks with Transformers, recently inserted on state of the art. Not limited by the comparison of model behaviors, it is possible to raise case studies on feature selection and time resolution, intended to see the impact of those features and the number of samples on the studied model’s architecture. Even though it is possible to evaluate the performance, the comparison becomes more complete with the interpretability of the models, noticing the feature force and average importance for the whole testing set. The comparison shows a great performance using recurrent models, Transformers, and those with convolutional layers. The case study shows that with the decrease of samples parallel with the resolution, the feature selection makes the accuracy degrade, thus, the selection makes good results where the training set comes with greater time resolution. All the results are possible to become better with the voting ensemble method, optimized by the simulated annealing metaheuristic.