Predicting energy poverty with combinations of remote-sensing and socioeconomic survey data in India: Evidence from machine learning

Highlights

• Remote sensing is combined with socioeconomic survey in energy poverty prediction.

• Machine learning algorithm is used for prediction of energy poverty in India.

• Geographical and environmental indicators account for 90.91% prediction power.

• Precipitation and PM2.5 are the two most significant predictors.

Abstract

Identifying energy poverty and targeting interventions require up-to-date and comprehensive survey data, which are expensive, time-consuming, and difficult to conduct, especially in rural areas of developing countries. This paper examined the potential of satellite remote sensing data in energy poverty prediction combined with socioeconomic survey data in response to these challenges. We found that a machine learning algorithm incorporating geographical and environmental remotely collected indicators could identify 90.91% of the districts with high energy poverty and performs better than those using socioeconomic indicators only. Specifically, precipitation and fine particulate matter (PM_2.5) offer the most significant contribution. Moreover, the algorithm, which was trained using a dataset from 2015, could also perform well to predict energy poverty using two environment indicators: precipitation and PM_2.5 concentration.

Introduction

Lack of access to modern energy has been recognized as a major hindrance to decent living standards, good health, education, and economic development. This paper aims to improve the identification of districts with a high prevalence of energy poverty to support policy efforts and better target interventions. Indeed, many governments worldwide have put programs in place to alleviate energy poverty. In India, where over 200 million people did not have access to electricity in 2016, a succession of programs, including the Deen Dayal Upadhyaya Gram Jyoti Yojana (DDUGJY) and the Saubhagya scheme, achieved great improvement in the number of households with access to electricity and put India on track to reach full electrification before 2030 (IEA, 2017). However, supply reliability and affordability issues mean that access to electricity is still a challenge for many and the gap to full electrification varies geographically, especially within states (Jain et al., 2018). In 2017, about 60% of the population still relied on biomass for cooking and averaged over one hour per day collecting fuel (IEA, 2017). The 2016 government program, Pradhan Mantri Ujjwala Yojana (PMUY), achieved notable success in shifting household consumption away from biomass toward increasing use of liquefied petroleum gas (LPG). However, here again, LPG adoption varies from state to state and within states (Jain et al., 2018).

One of the reasons it is difficult to tackle energy poverty lies in its multiple causes and repercussions. In developing countries, energy poverty describes insufficient access to modern energy services: lighting and powering appliances via electric power, and clean cooking via LPG (Li et al., 2014). Specifically, the reliance on traditional biomass (dung cake, agro residues, firewood), and to a lesser extent on kerosene, is problematic due to the unfair burden it places on energy-poor households with respect to fuel collection, fuel inefficiency, health consequences, or financial implications (Heltberg, 2003; Malla and Timilsina, 2014; Poblete-Cazenave and Pachauri, 2018). Day et al.'s (2016) definition, based on Sen's capability approach, is particularly suitable for capturing the multidimensional nature of energy poverty:

An inability to realize essential capabilities as a direct or indirect result of insufficient access to affordable, reliable and safe energy services, and taking into account available reasonable alternative means of realizing these capabilities (Day et al., 2016).

This approach is also consistent with recent findings from high-income countries that energy poverty is associated with decreased subjective wellbeing (Churchill et al., 2020; Welsch and Biermann, 2017). Estimations of energy poverty have moved from simple, economic-based measures to a range of multidimensional energy poverty indexes (MEPI) that focus on physical access to electricity, affordability, and access to equipment, cost, and efficiency.

The many factors involved in measuring energy poverty with a MEPI require information on socioeconomic indicators and energy-specific information such as equipment, connection, price, and availability of electricity and fuel. Accordingly, an important obstacle to measuring energy poverty is data availability. This is especially true in rural areas of developing countries where practical access to the areas to conduct field surveys and the financial means to conduct the surveys are often challenging. The difficultyto collect data on a wide scale is such that institutional measures and reports are often restricted to the concept of electricity connection and omit the other aspects of energy poverty (e.g., affordability, reliability, safety, availability).

In this regard, data collected via satellite remote sensing (RS) techniques offer promising possibilities to support data collection efforts. Thanks to their wider availability and improvement in quality and coverage, RS data have found applications in many fields and have already proven to be useful to predict income poverty, alone or in combination with field data. They offer inexpensive information, comprehensive area coverage, and frequent updates. Although they cannot replace detailed individual surveys conducted on the field, they can be employed to identify areas where risk is higher. This paper observed how an energy poverty index, built with field survey data from the Council on Energy, Environment and Water (CEEW) in India, can be predicted using RS data and standard socioeconomic information. Using a random forest algorithm, it shows that RS data are good predictors of energy poverty prevalence. Thus, it can produce cost-effective information for large geographic areas to better outline risk areas and target policy attention to aid energy poverty alleviation.

In the process, we questioned the relationship between energy poverty and the physical environment surrounding households and communities. Socioeconomic causes and consequences of energy poverty have been largely studied, with income and affordability recognized as the main determinants of energy poverty. However, there has been little examination of the impact of natural elements, except for outdoor temperatures, on energy vulnerability. Thus, the first contribution of this paper is to look at whether geographical and environmental factors can be used to predict energy poverty. The use of RS data to highlight the geographical and environmental aspects of energy poverty has received limited attention, despite some studies showing that geographic information system (GIS) analysis is particularly well-suited for energy poverty prediction. The second contribution is the use of a machine learning approach to predict the districts with a high prevalence of energy poverty. Although the existing literature has substantiated the outperformance of machine learning in terms of energy prices forecasting (Herrera et al., 2019; An et al., 2019), to the best of our knowledge, very little attention has been paid to the prediction of energy poverty. To fill this research gap, we used the random forest model, a prevalent machine learning algorithm, to predict energy poverty. This paper focused on India because a large percentage of the world's energy poverty comes from India (Bhide and Monroy, 2011).