Comparative Analysis of Energy Poverty Prediction Models Using Machine Learning Algorithms

• Published in: Journal of Korea Planning Association Vol. 56; no. 5; pp. 239 - 255
• Main Authors: Hong, Zhe, Park, In Kwon
• Format: Journal Article
• Language: English
• Published: 대한국토·도시계획학회 31.10.2021
• Subjects: 공학일반; 예측모형; Nonlinear Relationship; 에너지 빈곤; Energy Poverty; 머신러닝 알고리즘; Machine Learning Algorithms; 비선형 관계; Prediction Models; Random Forest; 랜덤 포레스트
• ISSN: 1226-7147; 2383-9171
• DOI: 10.17208/jkpa.2021.10.56.5.239

The energy poor are vulnerable to energy costs, and their vulnerability is expected to increase with climate change. Therefore, accurately predicting energy poverty can help minimize damage cause by climate change to achieve social equity. This study aims to develop a series of models to predict energy poverty as well as analyze the relative importances and partial dependences of predictors by applying machine learning algorithms. Accordingly, we used the 2016 Household Income and Expenditure Survey data and applied different machine-learning methods, such as Decision Tree, Artificial Neural Network, Bagging, Random Forest, Extreme Gradient Boosting, and Support Vector Machine. The main results are as follows: First, the Random Forest model performs the best at predicting energy poverty. Second, household income, food expenses, living floor area, age of the householder, public transfer income, and educational attainment of the householder are the most important predictors. Third, using partial dependence plots (PDPs) and accumulated local effects (ALEs), we identified the nonlinear relationships between the six most important predictors and the response variable. Based on these findings, we expect to derive meaningful policy implications to identify the traits that affect the probability of a family’s energy poverty. KCI Citation Count: 0