Previsão de séries temporais de consumo de energia elétrica com aprendizagem profunda para um sistema IoT

Abstract

Energy consumption and energy efficiency are topics that have attracted the attention of researchers in recent years, in order to seek scientific and technological solutions for energy production and cost reduction. One of the alternatives that have obtained satisfactory results is the use of technologies based on Internet of Things (IoT) and Deep Learning (DL). Based on this, it is proposed to evaluate the performance of Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) Deep Neural Networks (DNNs) for the prediction of univariate Time Series (STs) of electricity consumption. Cross-validation for Time Series (CV-TS) was used for this purpose. The results indicated that the LSTM models tended to perform better in comparison with the Extreme Gradient Boost (XGBoost) and Random Forest (RF) algorithms, and that in turn the BiLSTM models performed better than the LSTM models, with a statistically significant difference according to Friedman's tests (p = 0.0455) considering four datasets. Thus, the comparative experimental results and statistical analyses corroborate that AP can be used for prediction obtaining better results, and that despite having a longer training time, BiLSTM was statistically superior to LSTM. Finally, it can be emphasized that one of the main advantages of prediction is the possibility, based on the integration of DL with IoT systems, to design effective short, medium and long term strategies to promote appropriate solutions for each situation.