Identifying at-risk populations is essential for designing effective energy poverty interventions. Using data from the HILDA Survey, a longitudinal dataset representative of the Australian population, and a multidimensional index of energy poverty, we develop a machine learning model combined with SHAP (SHapley Additive exPlanations) values to document the short- and long-term effects of individual and contextual factors—such as income, energy prices, and regional conditions—on future energy poverty outcomes. The findings emphasize the importance of policies focused on income stability and may be used to shift the policy focus from reactive measures, which address existing poverty, to preventive strategies that target households showing early signs of vulnerability.