Energy consumption level prediction based on classification approach with machine learning technique

Abstract

Most of researches primarily use regression-type prediction, a method of estimating a numerical value given its historical data. In this study, a novel and practical prediction technique based on predicting identified energy consumption levels (e.g. low, average, and high levels) is proposed as an alternative on this conventional regression approach. Temperature and time features serve as predictors for short-term energy level prediction using commonly-used machine learning classifiers - artificial neural networks, support vector and random forest. The energy consumption numerical values were classified into ordinal bins created using a general percentile statistic. For verification, training subset underwent 10-times cross validation and test models are verified using the testing subset. These splitting and validation process are repeated 10 times with random permutation in each run prior to splitting into train and test subsets. The three-level energy prediction results show at least 90 % classification accuracy using any of the classifiers. However, the higher the number of desired energy levels, prediction accuracy tends to decrease. On the other hand, misclassification of energy level tends not to deviate from its true level that would still give confidence on the prediction performance when applied to any energy management system.