Accurate forecasting of household waste generation is essential for sustainable urban planning and the development of data-driven environmental policies. Conventional statistical models, while simple and interpretable, often fail to capture the nonlinear and multidimensional relationships inherent in waste production patterns. This study proposes a machine learning-based regression framework utilizing Random Forest and XGBoost algorithms to predict annual household waste generation across four metropolitan regions in South Korea Seoul, Gyeonggi, Incheon, and Jeju over the period from 2000 to 2023. Independent variables include demographic indicators (total population, working-age population, elderly population), economic indicators (Gross Regional Domestic Product), and regional identifiers encoded using One-Hot Encoding. A derived feature, elderly ratio, was introduced to reflect population aging. Model performance was evaluated using R2, RMSE, and MAE, with artificial noise added to simulate uncertainty. Random Forest demonstrated superior generalization and robustness to data irregularities, especially in data-scarce regions like Jeju. SHAP-based interpretability analysis revealed total population and GRDP as the most influential features. The findings underscore the importance of incorporating economic indicators in waste forecasting models, as demographic variables alone were insufficient for explaining waste dynamics. This approach provides valuable insights for policymakers and supports the development of adaptive, region-specific strategies for waste reduction and infrastructure investment.
Lee et al. (2025) studied this question.