Energy poverty prediction and effective targeting for just transitions with machine learning

The prevalence of energy poverty as a major challenge in numerous countries, the escalating energy crisis and the need to build just supporting mechanisms within the net zero energy transition add impetus to improving our ability to accurately predict energy vulnerable households. In Europe, this is hindered by limited recognition of the fact that energy vulnerable households are not necessarily income poor (and vice versa). Artificial Intelligence, and machine learning techniques in particular, may be applied to improve the targeting mechanism of energy poverty schemes, enabling accurate prediction of energy vulnerable households via objective, publicly available data. However, such applications are still limited, especially across a large number of countries. In response to the above, we develop an innovative machine learning framework for accurate prediction and fair targeting of energy poor households across all the current members of the European Union, and the United Kingdom. While we explore various machine learning algorithms, most of our analysis is performed using a Random Forest classifier. Our approach to explore energy poverty beyond income reveals household-level and country-level predictors of energy poverty, such as dwelling condition, energy efficiency, social protection payments and gas supplier switching rates. We also demonstrate how machine learning algorithms offer straightforward visualization of the mechanism that determines the energy poor classification, improving the transparency of alleviation schemes and assisting policy-makers in setting effective thresholds for assistance allocation. Finally, we evaluate the potential fairness of alleviation schemes and demonstrate that basing their targeting exclusively on income-relevant or social welfare-relevant criteria would be ineffective and result in significant numbers of energy poor households being excluded from energy assistance.