Data-driven analysis for the characterization of household appliance ownership and use in Sub-Saharan Africa

Data-driven analysis for the characterization of household appliance ownership and use in Sub-Saharan Africa Nicol` o Stevanato a,* , Davide Fioriti b , Tommaso Ferrucci a , Luca Belloni a , Lorenzo Rinaldi a , Davide Poli b , Emanuela Colombo a a Department of Energy, Politecnico di Milano, Milan, Italy b DESTEC, Universit` a di Pisa, Pisa, Italy ARTICLE INFO Keywords: Access to electricity Load estimation Appliance adoption Data analysis Electricity demand drivers ABSTRACT To grant reliable and affordable electricity provision to non-electrified communities, proper system sizing, based on accurate demand estimation is crucial. However, the absence of historical data, and scarce, scattered, and often unreliable pre-electrification surveys, make this process particularly prone to errors. Acquiring data, especially with high quality and detail, is difficult, time-consuming and expensive. Even though, in a few sitespecific cases the limited data collected has allowed researchers to develop methodologies to generate synthetic demand profiles based on variegated site-specific socio-economic information and appliance adoption patterns. However, given the lack of comprehensive datasets of such information, the use of synthetic methodologies has been circumscribed to limited regional and socio-economic scopes. This research proposes the development of a data-driven machine-learning framework for estimating appliance adoption patterns with a subset of relevant socio-economic indicators, identified throughout a comprehensive literature analysis and data collection across various sources. To successfully train the model, a novel open-access database has been created and populated with socio-economic information combined with appliance data collected from public and private sources. Finally, a structured logistic regression analysis has been performed, not only to capture the nexus of socio-economic factors with appliance adoption but also to estimate the most relevant ones. The methodology calibrated with the proposed open-access database has shown 71.7 % accuracy, which represents an important achievement in the field. The study’s findings lay the foundations for simplifying the estimation of appliance adoption, which can facilitate the demand estimation for sizing rural energy systems and rural electrification approaches. Table of Abbreviations CEFA Comitato Europeo per la Formazione Agraria CLASP Collaborative Labelling and Appliance Standards Program CV cross validation DC developing country D ed C developed country ESMAP Energy Sector Management Assistance Program FN false negative FP false positive FUNAE Fundo de Energia HH household HHH household’s head LSMS+Living Standards Measurement Study –Plus NaN not a number (continued on next column) (continued) CEFA Comitato Europeo per la Formazione Agraria PV photovoltaic SDG Sustainable Development Goal SHS solar home system SME small medium enterprise SSA Sub Saharan Africa TN true tegative TP true positive 1. Introduction Energy plays a crucial role in addressing modern global challenges, * Corresponding author. E-mail address: [email protected] (N. Stevanato). Contents lists available at ScienceDirect Renewable and Sustainable Energy Reviews journal homepage: www.elsevier.com/locate/rser https://doi.org/10.1016/j.rser.2025.115396 Received 22 November 2022; Received in revised form 12 December 2024; Accepted 17 January 2025 Renewable and Sustainable Energy Reviews 212 (2025) 115396 1364-0321/© 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). as emphasized by the 7th Sustainable Development Goal (SDG) of the United Nations in the Agenda2030. It enables local development [1,2] by facilitating the creation of income-generating activities, which can reinforce growth through a positive feedback loop. However, energy is not the sole determinant; socio-economic, infrastructural and cultural factors also exert a strong influence [3–5]. Despite its priority, achieving universal access by 2030 under SDG7 remains out of reach [6], a situation exacerbated by Covid-19 pandemic, which slowed the process [7]. Access to electricity is lower in rural areas than in urban ones [8]. The centralized demand of urban areas justifies the investment in expensive infrastructure, such as national grid extension. Conversely, scattered and scarcely populated areas, such as the rural ones in Sub-Saharan Africa (SSA) [9], are often better served by decentralized energy sources, which require lower initial investments and can be tailored to local needs. Data challenges also pose a barrier, as information is often scattered across multiple sources, inconsistent in format, and limited in quantity or quality [10]. This hinders the profitability of projects and the deployment of solutions to improve the conditions of the average SSA household, which today consumes less than 20 % of the energy consumed by households in Developed Countries (D ed Cs) [11]. The financial viability of decentralized energy projects depends on accurate planning to meet community needs, or demand. Oversized designs hinder cost recovery [12], and increase tariffs and upfront costs [13,14]. Undersized systems may slow down local development by failing to meet demand growth [15] or by providing unreliable supply, limiting growth [16]. Demand assessment has traditionally relied on pre-electrification interviews, which are often inaccurate [17,18]. Alternative modern approaches use computer models, for planning [19] and load prediction [20]. Bottom-up models offer high accuracy in demand estimation [21, 22] and long-term trends identification [23], through data-driven approaches that capture local dynamics. However, their scalability remains uncertain, due to limited data for calibration and validation, including appliance adoption information [19]. Efforts to address data issues through regrouping and harmonization have shown limited success [17]. Data availability remains a significant challenge, as it is often fragmented, inconsistent, or of poor quality [10]. This data scarcity is a significant barrier to deployment of energy access projects that this study seeks to address. This study proposes a data-driven machine-learning framework to estimate appliance adoption based on socio-economic and geographical factors across Sub-Saharan countries. A comprehensive literature review identifies key determinants of energy demand, while a data harmonization process creates a novel open database to train the model. Major socio-economic and geographical factors are identified through the review, and relevant data sources are collected and harmonized. Logistic regression models are then developed to estimate appliance adoption, highlighting the most relevant input factors for prediction. This study introduces several novelties: the development of datadriven machine-learning models to estimate appliance adoption with a regional scale for electrification planning, the creation and harmonization of a database of socio-economic and geographical indicators appliance adoption estimation, the identification of key input factors to improve prediction accuracy and enable simplified models, and the calibration and validation of the proposed comprehensive procedure. Fig. 1 provides and overview of the work, detailing the three steps leading to the final goal: constructing a prediction algorithm trained on a database of appliance adoption patterns and socio-economic conditions, with explicit references to the sections. 2. Literature review on rural communities load demand estimation A review of the scientific literature is conducted to identify the main drivers influencing community electricity demand, forming the foundation for constructing the appliance ownership database. To accelerate Fig. 1. Structure of the three steps that compose the activity of the work. N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 2 the development of energy projects, priority was given to studies that provide inputs for Energy Planning Models. The review focused on energy access research within the last 15 years, with a focus on drivers impacting electricity consumption. A systematic search was performed using Scopus with keywords: ("electricity demand" OR "load profile" OR "appliance ownership" OR "demand evolution" OR “electricity consumption patterns”OR “determinants”OR “driver”). Relevant studies were identified through a manual examination of the selected works that mentioned energy demand determinants, which findings are discussed in this section. The review has two objectives: to identify drivers previously used to explain variations in load profiles, energy consumption, and appliance adoption, and to establish a starting point for this study, by summarizing state of the art of the methodologies for rural areas demand estimation, which this work aims to refine and build upon. A total of 39 works, including five reviews, were identified and analysed. While the focus was on studies regarding DCs, selected applications to D ed Cs were included when their methodology were relevant and applicable to DCs. As pointed out by Debnath and Mourshed [24], the majority of current Energy Planning Models were created in D ed Cs and often rely on assumptions that that fail to address the unique challenges of DCs. These challenges include meeting basic socio-economic needs and estimating suppressed demand, which are taken for granted in D ed Cs but requires tailored solutions in DCs. Similarly, demand estimation for un-electrified areas must account for households that may have no prior experienced with electricity, necessitating customized approaches. Research efforts have been carried out to fill the gap. In their review of energy planning case studies in remote areas of DCs, Riva et al. [19] proposed a classification in terms of spatial coverage, planning horizon, energy carrier, decision criteria, mathematical model, and demand sector. They found that spatial coverage influenced the selected approach. Large-scale planning at regional or national levels often relied on simplistic, top-down methods using aggregated data to estimate demand and its evolution over time. In contrast, small-scale systems planning employed more complex, bottom up, models that utilized on-field data, better capturing local socio-economic and cultural dynamics. Data scarcity, however, often limited these models, forcing them to adopt simplified assumption, such as fixed demand over time. Kuster et al. [20] reviewed 113 electrical load forecasting models, emphasizing the intensive data requirements of bottom-up approaches for long-term demand prediction. Their study classified model input variables into socio-economic, environmental (e.g.: weather conditions), building and occupancy (e.g.: dwelling characteristics), and Table 1 Classes of drivers identified in literature. Publication Socio-economic Dwelling Appliance Past demand Supply Alternative energy sources Geographical Cultural Literature with focus on Appliance Adoption [29]    X   [26] X X X X  X [27] X X X   X X X [28] X X   X X Literature with focus on Aggregated Electricity Demand [31] X X    X  [32] X    X   [17] X        [33] X    X X   Literature with focus on Load Profile [39]   X    [22] X X     [42]   X X    [18]  X     [21] X X     [43] X   X X X [37] X X X    X X [40] X X X X X    [41] X X    X [38]   X  X [36] X X X X   X [35] X X X X X    Literature with focus on Demand Evolution [44] X    X   [1] X X X X X X X [45] X X X X X X X [47] X X X X X X X Literature with Multiple Foci [48] X X    X X  [49] X X     X [51] X X X X X   [55] X   X    [56] X    X   [50] X X X      [53] X        [54] X X X X X X  [52] X X X   XX [57] X   X    Total 29 12 17 13 14 10 11 9 N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 3 time-index (e.g.: past demand data). Jones et al. [25] categorized determinants of electric consumption into socio-economic, dwelling and appliance factors. This classification has been adopted and expanded in this work to include additional drivers identified in the literature review. The final classification of drivers encompasses socio-economic, dwelling, appliance, past-demand, supply, alternative energy sources, geographical and cultural factors. Socio-economic drivers broadly cover social and economic conditions at both community and individual levels, such as population density, household income and composition, and business revenues. Dwelling drivers relate to household characteristics like the number of rooms, while appliance drivers focus on factors such as price and nominal power. Past-demand data includes modelling techniques relying on historical electricity usage, such as past load profiles. Supply drivers address the supply side of the electric system, including the number of hours of electricity availability. Drivers related to alternative energy sources consider factors such as the price of kerosene or candles. Geographical drivers account for location-specific characteristics, such as climate zones and distance from the nearest city. Cultural drivers involve societal habits and practices. The analysed publications are categorized by their focus - appliance ownership, aggregated electricity demand, load profile, demand evolution over time, and multiple foci. Each study is briefly presented to derive insights into the methodologies and drivers relevant to this work. Table 1 summarizes the factor classes and their corresponding drivers as identified in the literature. 2.1. Focus - appliance ownership Regression techniques were employed in all studies modelling appliance ownership. Rao and Ummel [26] employed logistic regression and boosted regression trees to predict ownership of refrigerators, washing machines and television. two covariates sets were assessed: a sparse set (income and urbanization) and a diverse set (socio-economic and cultural factors). The results showed that the input variable set had a greater impact on prediction accuracy than the choice of the model. Kurata et al. [27] used ordinary least squares regression on field survey data in Bangladesh to predict solar home systems (SHS) ownership, differentiating residential and business users. They found business users more sensitive to energy costs. Similarly, Richmond et al. [28] applied ordinary least square models to data from 5000 Indian households, analysing the influence of electrification duration on appliance ownership. Ownership was evaluated by single appliances, ownership tiers, and the number of appliances per type. Time since electrification positively influenced all outcomes, emphasizing the need to account for time dynamics in system sizing. A survey in Western Kenya [29] examined the relationship between household connection type (national grid, SHS, or no connection) and appliance ownership. SHSs influenced ownership of phone chargers and televisions, but had limited impact on other appliances. SHS users were generally of higher socioeconomic statuses compared to households relying on kerosene or without electricity. As Lukuyu [30] noted, SHS represent a key step in the energy ladder, improving electricity access, and enabling higher appliance ownership through economic development. 2.2. Focus - aggregated electricity demand Louw et al. [31] estimated average electricity demand of two rural villages in South Africa by collating two survey-based datasets. Using a log-linear regression model, tested the significance of various input variables and identified income as a key driver, concluding that the use of electricity is cost-sensitive. Azadeh et al. [32] used regression alongside an artificial neural network to estimate annual household electricity consumption in Iran. Dominguez et al. [33] applied linear regression to analyse electricity demand and its main drivers in rural Kenya. Blodgett et al. [17] compared metered demand data with survey-based bottom-up estimates for eight Kenyan communities. Their findings revealed that survey-based methods overestimated demand by 330 %, potentially leading to substantial sizing error. The authors proposed a proxy methodology, supported by an analysis of variance (ANOVA) test, which demonstrated that consumers from different communities belonged to the same sample population. This approach allowed consumption data to be interchanged, reducing estimation errors without requiring distance metrics, thereby simplifying the analysis and focusing on available data. 2.3. Focus - load profile Hartvigsson and Ahlgren [18] conducted a survey-based demand estimation in a Tanzanian village to assess its accuracy in predicting load profiles. Despite the village having electricity for over 10 years, survey responses underestimated actual metered usage. While traditional bottom-up, survey-based approaches often lead to poor system sizing, innovative methods in the literature show promising results. LoadProGen [22] and RAMP [21] are stochastic load profile generators that use the same data pool of traditional approaches, such as planned appliances and usage patterns. However, Takalani et al. [34] highlighted the unreliability of surveyed data on planned appliances purchases, underscoring the need for methodologies that estimate appliance adoption using socio-economic and weather data. RAMP has also been integrated into the M-LED platform [35], an demand estimation tool that combines multi-sectorial, granular data to generate monthly load curves for rural communities. Adeoye and Spataru [36] simulated and forecasted hourly and yearly power consumption for 14 West African countries for 2016 and 2030 using a hybrid demand model combining top-down and bottom-up approaches. The model accounted for nine household appliances types, occupancy patterns, weather, day type, and daylight hours to estimate hourly electricity usage in urban and rural electrified households. By incorporating a random generator, national survey occupancy data, and average appliance operation timings, the model calculated the synthetic likelihood of appliance usage. Caquilpan et al. [37] developed a novel method to estimate home load profiles by combining electrical power consumption and sociodemographic data. Self-organizing maps, a neural network, classified homes based on unique characteristics, while a probabilistic model using Bayesian networks simulated appliance behaviour, incorporating daily demand variations. Few et al. [38],applied an open-source framework for microgrid design to estimate load demand, optimize cost, and characterize systems. Their study evaluated the influence of weather, rurality and non-domestic demand on microgrid demand profiles. They found that weather impacts, particularly in hot regions, increase with higher levels of energy access, with cooling technologies such as fan and refrigerators affecting nighttime demand. Hernandez et al. [39] demonstrated the application of machine learning for short-term microgrid load forecasting, using only two years of past consumption data and a consumption calendar. Similarly Dominguez et al. [40] used supervised and unsupervised models to estimate hourly lighting load profiles in rural Kenyan and Tanzanian households. Their models integrated publicly available data at multiple scales with satellite imagery. Scott and Coley [41] analysed metered data from two Tanzanian solar PV diesel hybrid mini-grids, categorizing homes are based on demand attributes. By integrating load profile data with household demographics and appliance ownership, they concluded that socioeconomic status, occupancy, and appliance ownership shape load profiles. Using data from eleven East African microgrids over two years, Williams et al. [42] examined weekly and monthly seasonality in consumption and long-term growth patterns. While most systems showed N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 4 demand growth, no universal trend emerged. Additionally, customers exhibited a preference for frequent, small payments, when given flexible tariff options, highlighting the importance of considering payment frequency in tariff models. Lorenzoni et al. [43] developed a database of load profiles from sixty-one mini-grid projects in DCs. A clustering identified archetypal profiles, followed by graphical analysis to explore factors influencing load profile shapes. 2.4. Focus - demand evolution Fobi et al. [44] conducted a longitudinal study of Kenyan residential electricity consumption, to explore demand evolution over time, addressing a gap identified in Ref. [19]. Their findings showed an overall increase in consumption, with urban households experiencing higher growth than rural ones. This underscores the importance of including urban-rural distinctions in future studies. Riva et al. [45,46] developed and calibrated a System Dynamics model to examine the relationship between socio-economic variables and electricity demand in a Tanzanian rural community. The model, based on causal-loop diagrams from earlier work [1], captures a positive feedback loop between electricity provision and community socio-economic characteristics. It was further tested in a subsequent study [47]. 2.5. Multiple foci Most reviewed works cover multiple focuses. Van Ruijven et al. [48] critiqued global energy models for excessive aggregation, building a bottom-up model to estimate household energy use, including cooking, water heating, space heating needs, appliance ownership, and lighting. Inputs included household (e.g., expenditure) and community (e.g., population) variables. Daioglou et al. [49] extended this model to five DCs, whereas the original analysis focused on India. Riva et al. [50] further built on Van Ruijven’s model combining load estimation, load profile generation and system sizing for an Indian rural community. Fabini et al. [51], used k-nearest neighbours’regression to predict the increase in appliance ownership expected after electrification. Socio-economic metrics, were employedto assess proximity between Kenyan wards, with ownership in electrified wards serving as a basis for predictions in unelectrified ones. These ownerships estimates were then converted into electricity demand using typical appliance consumption values. As Fabini et al. state: “Implicit in this approach is the assumption that localities that share socioeconomic characteristics will also have similar demand for electricity services and similar ability to pay for them.”This methodology aligns with the proxy approach from Ref. [17], while introducing multidimensional distance metrics. However, the limited focus on socio-economic features could reduce predictive accuracy, suggesting that expanding the range of input features could enhance results. Allee et al. [52] also employed machine learning models, emphasizing the value of on-field data for proxy based approaches, despite the poor reliability of survey based appliance ownership data. They performed variable importance tests to reduce data collection needs, offering potential time and cost savings. In contrast, Shibano et al. [53] developed an income-based model to estimate electricity consumption, linking appliance ownership (modelled with a Gompertz curve) to consumption using a Gamma distribution calibrated via regression. Poblete-Cazenave and Pacahuri [54] used micro-data from national surveys to estimate electricity demand in developing nations. They found that appliance adoption varies with nation, appliance type, environment, and income. In all four studied countries, entertainment accounted for a significant share of consumption, while rising incomes shifted demand toward food preparation and preservation, and clothing care. Williams et al. [55] conducted a customer-focused analysis, segmenting 821 using a k-means clustering algorithm based on mean normalized load curves and daily electricity consumption. Exploratory graphical analysis sought to link these variables to customer categories, such as public premise, home, business, or mixed-use. Bahaj and James [56] compared load profiles and daily electricity demand across four rural microgrids with the same capacity. Consistent with Louw et al. [31], they found that consumption dynamics are cost-sensitive, driven by the relationship between income and tariff. In one system, prepayment adoption reduced overdue payments and dropouts, increasing consumption. Hartvigsson et al. [57], analysed a Tanzanian microgrid at shortand long-term scales. Load profiles were evaluated for performance metrics Fig. 2. Structure of the proposed database. HH: Household. N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 5 like load factor, while 30 months of electricity expenditures, tied to tariffs, trained regression algorithms. Results revealed strong seasonality in consumption due to the agricultural-based economy. Business load profiles also varied significantly by type emphasizing the need to consider business activities and seasonality in system sizing. Socio-economic drivers are the most commonly used across the reviewed literature, followed by drivers related to the techno-economic parameters of appliances and electricity supply technologies. Cultural drivers are the least considered, likely due to the challenges of data collection and classification, as well as the engineering focus that excludes such analysis [58]. In both appliance ownership and load profile literature, most driver classes are represented. However, within appliance ownership studies, this broad coverage is consistent across individual studies (except [29]). whereas load profile studies often focus on a narrower set of categories. Similarly, studies on demand evolution show near-complete coverage of driver categories. Given the goal of this research –and the dependence of electricity demand, load profiles and demand evolution on appliance ownership and usage patterns –the following sections will address all the identified drivers as drivers as influencing appliance adoption and use, forming the basis for the proposed methodology. 3. Data preparation and structure This section describes the data collection and cleaning procedure used to build the dataset for training the machine-learning model discussed in Section 4. 3.1. Data collection The dataset was built through an initial data collection covering most factors reviewed in Section 2, followed by the harmonization and integration of diverse data sources. Fig. 2 depicts the final structure of the harmonized database, where columns represent the drivers and appliance ownership data, and rows correspond to entries from different data sources, villages, or households. This structure streamlined data integration, as additional sources simply added rows without altering the column structure. Three actions were undertaken to populate the dataset: i) interviews with energy access sector stakeholders to obtain data on appliance adoption and household socio-economic characteristics, ii) collection of existing databases relative to recently electrified communities based on Fig. 3. Flowchart of Database filling. Sources: Beck et al. [64]; World Bank [65]; JRC [66]; NASA [67]; GPW v4: Gridded Population of the World version 4 [68]; Meta - Data for Good [69]. QGIS logo under CC BY-SA 3.0 license. Table 2 Share of available drivers and number of respondents for each source. Source Share of drivers N◦of respondents CEFA 57 % 254 CLASP 23 % 2729 ESMAP 71 % 11875 FUNAE 24 % 210 LSMS+64 % 1184 Fig. 4. Response rate duration curves. N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 6 Table 3 Resume table of appliance adoption data in each source. Red: the variable is not present in the source. Yellow: the variable is present only in some of the source’s entries. Green: the variable is present in all source’s entries. Variable ESMAP CLASP LSMS+ CEFA FUNAE Fig. 5. Flowchart for prediction model development. Fig. 6. Appliance presence recurrency within the database (in red the eight appliances considered in the analysis, being the most recurrent). (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.) N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 7 field interviews, and iii) extraction of information from geographically explicit datasets to enrich the database. 3.1.1. Interviews with energy access stakeholders Energy sector stakeholders in SSA, including private and public entities, academia, NGOs and international organizations, were contacted to share data from their past electrification projects. The goal was to gather information on household appliance use and socio-economic conditions. These interaction also allowed for testing the completeness and significance of the driver’s list. In a few cases, key insights led to the inclusion of drivers not identified in the literature review. Three institutions shared on-field survey data. The Collaborative Labelling and Appliance Standards Program (CLASP), shared information from seven surveys conducted among customers of efficiency-based appliances. The Comitato Europeo per la Formazione Agraria (CEFA) contributed data from five surveys in Tanzanian communities electrified by microgrids, as well as grid-connected and not electrified communities. Fundo de Energia (FUNAE), shared information from three surveys conducted in communities electrified by their microgrids in Mozambique. 3.1.2. Collection of existing databases To further enhance the database, data from the World Bank’s opendata portal were integrated. Five datasets were selected and grouped into two categories. The first includes ESMAP campaigns for the Energy Access Diagnostic Reports based on the Multi-Tier Framework [59], with surveys from thirteen countries conducted between 2017 and 2019. Data from Kenya [60], Nigeria [61] and Zambia [62] were added to the database. The second category includes the Living Standards Measurement Study –Plus (LSMS+), which collects data on respondents’economic opportunities and welfare. From this programme the Tanzanian survey was included [63]. 3.1.3. Extraction of data from GIS databases To further enrich the databases georeferenced factors were collected for each entry from the referenced sources. These factors include climate zone, derived from a K¨ oppen-Geiger classification [64]; distance from the nearest national grid, based on World Bank data of built and planned African distribution lines [65]; distance from the nearest city, sourced from the Joint Research Centre’s (JRC) dataset of urban centres defined by population and built-up area thresholds [66]. Elevation data, measured in meters above sea level, obtained from the NASA’s Shuttle Radar Topography Mission (SRTM) [67]. Population density estimates are sourced from the Gridded Population of the World (GPW) dataset, aligned with national censuses [68]. The relative wealth index, developed by Data For Good, estimates living standards using connectivity data, satellite imagery, and additional sources [69]. To ensure consistency across countries, geospatial variables are harmonized using the Database of Global Administrative Areas (GADM), which provides a unified categorization of administrative divisions. The proposed data collection process harmonizes different datasets, enriching them with cross-referenced information suitable for machinelearning models. This approach is a novel contribution to appliance adoption studies and can be easily adopted and improved due to its open-source nature [10]. 3.2. Database structure Fig. 3 outlines the database construction process and provides an overview of its size and scope. The database includes sixty drivers of appliance adoption - potential factors influencing ownership –and seven appliance variables to be estimated by models trained on the drivers. Table 2 summarizes the total number of respondents and the percentage of input variables present per each source. ESMAP is the richest source, both in terms of completeness and respondents count. While CEFA and LSMS +have a high share of input variables, CEFA has fewer respondents. CLASP and FUNAE, have fewer input variables, making them less suitable for regression analysis due to the potential loss of predictive variables. Geographically, Kenya has the highest representation in the dataset with 6534 entries, followed by Nigeria (3669), Zambia (3622) and Tanzania (1680). Fewer instances are available for Uganda, Mozambique, Rwanda, and Senegal. This distribution reflects the geographic scope of the sources, as the three most represented countries are covered by ESMAP. Table 2 summarizes data availability across drivers by source, but it does not capture the completeness of individual variables. A variable inclusion in each source does not indicate how many non-empty entries it contains, as response rates vary within sources. Fig. 4 illustrates this Fig. 7. Database size against threshold in agreement with Eq. 1. Table 4 Appliance presence distribution. Appliance Presence Absence Delta Phone Charger 0.662 0.338 32 % Light bulb 0.865 0.135 73 % Iron 0.513 0.487 2 % TV 0.714 0.286 43 % Refrigerator/Freezer 0.437 0.563 13 % Radio/Stereo 0.515 0.485 3 % DVD Player 0.479 0.521 4 % Fan 0.425 0.575 15 % Table 5 Values of the confusion matrices of the unbalanced and balanced models. Model TP FP TN FN Accuracy Phone Charger Unbalanced 0.89 0.66 0.34 0.11 0.7 Phone Charger Balanced 0.69 0.42 0.58 0.31 0.65 Lightbulb Unbalanced 1 1 0 0 0.87 Lightbulb Balanced 0.65 0.41 0.59 0.35 0.64 Iron Unbalanced 0.75 0.29 0.71 0.25 0.73 Iron Balanced 0.74 0.27 0.73 0.26 0.74 TV Unbalanced 0.94 0.69 0.31 0.058 0.76 TV Balanced 0.71 0.32 0.68 0.29 0.7 Refrigerator/Freezer Unbalanced 0.74 0.17 0.83 0.26 0.79 Refrigerator/Freezer Balanced 0.79 0.21 0.79 0.21 0.79 Radio/Stereo Unbalanced 0.66 0.4 0.6 0.34 0.63 Radio/Stereo Balanced 0.64 0.37 0.63 0.36 0.64 DVD Unbalanced 0.62 0.34 0.66 0.38 0.64 DVD Balanced 0.67 0.39 0.61 0.33 0.64 Fan Unbalanced 0.73 0.22 0.78 0.27 0.76 Fan Balanced 0.82 0.3 0.7 0.18 0.75 N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 8 issue using source-specific response rate duration curves, which plot the number of variables (x-axis) with response rates equal to or exceeding the value on the y-axis. Fig. 4 indicates that ESMAP, LSMS+, and CEFA are the most suitable for statistical analysis due to their higher response rates compared to the other two sources. Analysing respondent categories reveals that 100 % of entries in ESMAP and LSMS +are households (as opposed to businesses or public services). Thus, using these sources for model training focuses the analysis to households. However, the methodology could be extended to non-households if sufficient data quality is ensured. Each database source is distinct in the appliances it investigates and the variables it collects. While appliance presence is the only adoption data common to all sources, other variables are more specific to individual datasets. Table 3 resumes this heterogeneity. Table 3 reveals partial coverage of output variables across most sources, with CLASP offering the least information, collecting only appliance presence. In contrast, CEFA is the most comprehensive, with all variables at least partially covered. For developing the machine learning models, three data subsets are used: “ESMAP”, from now referred to as “ESMAP original”,“LSMS+”, and “CEFA”. Fig. 8. Accuracy comparison between Logistic regression and Dummy classifier in “ESMAP original”. Fig. 9. Accuracy trends during Recursive Feature Reduction. Table 6 Comparison between the four final configurations. Configuration N◦input variables N◦respondents A oversampled 23 3106 A enlarged 32 2165 B oversampled 13 3738 B enlarged 25 2012 N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 9 [75] Bujang MA, Sa’at N, Sidik TMITAB, Joo LC. Sample size guidelines for logistic regression from observational studies with large population: emphasis on the accuracy between statistics and parameters based on real life clinical data. Malays J Med Sci 2018;25:122–30. https://doi.org/10.21315/mjms2018.25.4.12. [76] James G, Witten D, Hastie T, Tibshirani R. An introduction to statistical learning: with applications in R. New York, NY: Springer US; 2021. https://doi.org/ 10.1007/978-1-0716-1418-1. [77] Suthaharan S. Machine learning models and algorithms for big data classification, vol. 36. Boston, MA: Springer US; 2016. https://doi.org/10.1007/978-1-48997641-3. [78] Uddin MF. Addressing accuracy paradox using enhanched weighted performance metric in machine learning. In: 2019 sixth HCT information technology trends (ITT), ras Al khaimah, United Arab Emirates. IEEE; 2019. p. 319–24. https://doi. org/10.1109/ITT48889.2019.9075071. [79] Imputation Method - an overview | ScienceDirect Topics n.d. https://www.scienc edirect.com/topics/mathematics/imputation-method (accessed March 28, 2022). [80] Stevanato N, Sangiorgio I, Baldelli G, Farah A, Mereu R, Colombo E, et al. Guidelines for demand data collection for rural access to electricity projects. In: 2023 IEEE PES/IAS PowerAfrica. Marrakech, Morocco: IEEE; 2023. p. 1–5. https:// doi.org/10.1109/PowerAfrica57932.2023.10363265. N. Stevanato et al. Renewable and Sustainable Energy Reviews 212 (2025) 115396 16