A Comparative Analysis of Machine Learning Algorithms in Energy Poverty Prediction

Given the limited potential of conventional statistical models, machine learning (ML) techniques in the field of energy poverty have attracted growing interest, especially during the last five years. The present paper adds new insights to the existing literature by exploring the capacity of ML algorithms to successfully predict energy poverty, as defined by different indicators, for the case of the “Urban Region of Athens” in Greece. More specifically, five energy poverty indicators were predicted on the basis of socio-economic/technical variables through training different machine learning classifiers. The analysis showed that almost all classifiers managed to successfully predict three out of five energy poverty indicators with a remarkably good level of accuracy, i.e., 81–94% correct predictions of energy-poor households for the best models and an overall accuracy rate of over 94%. The most successful classifier in terms of energy poverty prediction proved to be the “Random Forest” classifier, closely followed by “Trees J48” and “Multilayer Perceptron” classifiers (decision tree and neural network approaches). The impressively high accuracy scores of the models confirmed that ML is a promising tool towards understanding energy poverty drivers and shaping appropriate energy policies.