Decision tree and reinforcement learning for contextual electricity consumption forecasting in buildings

A promising opportunity to optimize energy control and storage is the use of prediction. Several forecasting algorithms from the artificial intelligence area like the Neural Networks or from the machine learning field as K-Nearest Neighbors and XGBoost are recommended for prediction tasks involving the estimation of energy patterns ahead of time. Additionally, it is recommended to apply wisely the forecasting algorithm relying on unique contexts that define different target periods. The method in this paper targets the prediction of consumptions of a building scheduled for all periods of five minutes of a week:•with the support of K-Nearest Neighbors and Neural Networks•a decision tree identifies unique contexts according to rules that rely on all the patterns with energy and sensors data of a building for periods of five minutes•a Multiarmed Bandit algorithm gifted with reinforcement learning capabilities selects the algorithm more convenient for prediction tasks.The results and conclusions indicate that identification of contexts through decision rules results in higher confidence bounds while evaluating the most effective forecasting algorithm. The SMAPE forecasting errors obtained in the third context were 3.54% with KNN and 4.79% with ANN. The obtained SMAPE forecasting errors in the fourth context were 4.91% with KNN and 4.54% with ANN.