Machine learning: ‘New electricity’ enabling access

This article appeared in the July 2020 print edition of PV magazine. The article archive can be accessed here.

Machine learning is helping project owners to more effectively predict energy demand in unelectrified areas, which is critical for efficient electricity planning and minigrid sizing. Source: AMMP Technologies.

Andrew Ng, a leading educator in machine learning, has described the technology as “the new electricity” for its ability to transform modern society. The roughly 780 million people globally without electricity access would happily settle for just the “old” electricity. But a growing amount of data from smart meters, satellite imagery, and other sources has allowed machine learning to play a bigger role in advancing energy access.

Financing the unbanked is no small feat. But for solar suppliers, promising approaches are being deployed. Nithio, a credit-risk analytics platform, leverages socio-economic indicators – combined with customer transaction histories and other alternative data sources – to create high-resolution credit assessment models for pay-as-you-go (PAYG) solar home system companies. As described by Madeleine Gleave, chief data scientist at Nithio, “PAYG companies need capital to distribute their products out to customers, and accessing scalable debt financing remains a key bottleneck for the sector.”

Currently, lenders lack accurate insight into the lending risk at a customer level in PAYG companies’ portfolios, which often make up the largest part of the balance sheet. “PAYG companies struggle to assess the creditworthiness of their customers because their customers often have no formal credit history. Nithio seeks to fill this data gap using machine learning and geospatial data to provide credit risk assessments for traditionally unbanked households,” explained Gleave. This can help PAYG companies make better operational and underwriting decisions at the customer level, and help investors better assess risk at the portfolio level.

“Our intent in providing risk analytics tools isn’t just to cherry-pick the wealthiest customers,” said Gleave. “Universal energy access is a key part of our mission.” Nithio is leveraging its prediction models to also help governments and donors identify which households are least able to afford energy services at commercial rates from private companies and then helps design results-based financing programs and targeted grant assistance for energy access. “The goal is to use data to help companies and governments match the right type of products and financing for every household.”

Kyle Bradbury of Duke University is using satellite and drone imagery to automatically detect solar PV panels, in order to more effectively map electricity infrastructure in data-poor regions. Source: USAID.

Boosting bankability

Village Data Analytics (VIDA), a product of consultancy TFE Energy, uses satellite imagery and machine learning to improve the site selection process by identifying the most promising regions for minigrid development. “The big question developers are seeking to answer is: If you build a minigrid, will people buy your power? Being able to answer this question accurately is the holy grail for minigrid development,” explained Philippe Raisin, product manager at VIDA. “Our tool leverages machine learning to do a first-pass prioritization that will show minigrid developers the most promising sites.”

Minigrid developers can then do a more detailed on-the-ground survey in the villages recommended by VIDA, saving the developers time and money. The technology is also being used to help governments in African countries with their Covid-19 response by prioritizing rural health centers for electrification.

Hendrik Broering – chief product officer at AMMP Technologies, an energy data operations company – outlined two applications of machine learning that AMMP is now pursuing: the detection of anomalies in energy production data, and the identification of system components in a hybrid power system.

“Currently, we require our customers to provide us with a line-diagram explain- ing the topology of their power system, so we know what types of generation they have and how those assets are connected. We have started looking into training machine-learning models that could predict the system configuration based only on the data feed from our customers,” explained Broering.

Emily McAteer, chief executive officer at Odyssey Energy Solutions, an investment and asset management platform for distributed energy, described how her company’s work is improving mini-grid planning as well. “Through managing large-scale off-grid electrification programs on our platform, we are reaching a scale where enough minigrid projects have completed the full development cycle. We can now help our users assess how initial site surveys compare with the actual power utilization.” McAteer went on to say that “improving unit economics is critical for scaling the minigrid sector, and we are using machine learn- ing to more accurately predict electricity demand from on-the-ground surveys to improve minigrid sizing.”

Adaptive planning

Minigrid site selection can be viewed in the larger context of electrification planning, which combines grid extension and off-grid solutions like minigrids with solar home systems to electrify a country. A technical report by the World Bank, Reliable and Affordable Off-Grid Electricity Services for the Poor, describes how traditional approaches to electrification planning are being replaced by GIS-based approaches. The report notes that GIS-based plans can be rolled out at twice the speed and half the cost.

Stephen Lee, a Ph.D. candidate studying energy systems and machine learning at the Massachusetts Institute of Technology (MIT), advocates a more dynamic paradigm for electrification planning over traditional approaches. “The current approach to electrification planning often entails the compilation of giant five-year planning documents which can be obsolete as soon as they are published – if fundamental assumptions prove to be wrong,” Lee told pv magazine. He outlined the idea of adaptive electricity access planning, “where governments and planners assess uncertainty in their assumptions and continually update their plans as new information arises.”

Optimization-based electrification models can further amplify the benefits of GIS-based approaches and enable adaptive planning. According to Lee, “the MIT-Comillas Universal Electricity Access Lab’s Reference Electrification Model (REM) has been used to inform electrification plans for Rwanda and Mozambique. In a few hours, the model can compare thousands of candidate plans before presenting the lowest-cost designs to meet specified levels of demand.”

A major benefit of optimization-based models is that they can be rerun automatically when new information is available. “For example, if expectations about the costs of battery storage change, a planner can quickly modify REM’s inputs and have updated least-cost plans before the next workday,” said Lee.

While models like REM can be powerful tools for planners, they also require lots of high-resolution input data about electricity demand and supply. Accord- ing to Lee, “procuring such input data for electrification planning is rarely easy, and machine learning offers exciting opportunities to make accurate input data available for larger parts of the globe.”

Successful, continuous and adaptive planning requires planners to maintain their planning initiatives over a long timeframe and make investments in infrastructure and improved information. To succeed in this effort, it is often best for local governments and organizations to perform electrification modeling and planning, which means capacity building is an essential part of achieving universal energy access. 

Research by Christopher Elvidge at the Colorado School of Mines as part of the e-GUIDE initiative leverages nighttime satellite images and machine learning to measure power reliability. Source: Christopher Elvidge

Crucial capacity

Currently, many skills to perform machine learning and electrification modeling are less present in under-electrified countries. Nathan Williams, an assistant professor at the Rochester Institute of Technology and the executive director of the Electricity Growth and Use in Developing Economies (e-GUIDE) Initiative, shared the work of Carnegie Mellon University’s (CMU) Africa campus in Kigali, Rwanda. Williams describes how the CMU Africa campus has students from over 17 countries across Africa.

“In my personal view, it is important to develop the skills on the continent. Most of my students are African and understand the context on the ground, and if you are not from the region, you have blind spots. Having local talent is critical for the long-term success of the energy access sector”, explained Williams. To that end, the eGUIDE initiative provides funding for African students to intern with companies in the electricity sector. Many of the CMU Africa students who intern with energy companies are hired full-time and then serve as resources within those organizations for machine learning and data analysis.

A recent policy brief co-authored by Rob Fetter, a senior policy associate at Duke University’s Energy Access Project, outlines two proposals that could both address the shortage of data and help bridge the skills gap in the energy access sector: a global energy data commons and data competitions that provide selective access to more sensitive data.

A global energy data commons could serve as a central data repository and community around which companies, non- profits, and researchers could publish data on the energy sector. As Fetter described it, “the data commons is driven by and responsive to a community of users. More than just a data repository, the commons would allow users to be part of a community that would answer and respond to specific questions or challenges.”

Kyle Bradbury, the managing director of the Energy Data Analytics Lab at the Duke University Energy Initiative, is part of a team that recently won a National Science Foundation grant to create a plan for a data commons focused around risk and resilience in the energy sector, and they remain interested in creating such a commons for energy access.

As Fetter’s brief highlights, “a challenge to building a data commons is that entities with access to some of the most valuable data, such as energy consumption, may not want or be able to make that data public.” In response to this challenge, the policy brief lays out how the second prescription, data analysis competitions, could allow companies and organizations to selectively share data with teams that offer the most compelling ideas.

The competition is inspired in part by Orange Telecom’s “Data for Development Challenge,” where a call for proposals solicited ideas on a broad set of topics. One team of researchers was then granted access to anonymized mobile phone user data in Senegal, which they used to estimate demand for electricity and design a least-cost electrification plan.

Open competition offers a more level playing field that could partially offset structural inequalities in the academic research community. According to Fetter, “many researchers from low- and middle-income countries do not have the same opportunities to collaborate and gain access to data as some members of elite academic institutions. An open competition could create more opportunities for researchers from, say, Uganda to suggest clever ideas that incorporate insights they have, that researchers at some elite U.S. or European university might not have. The competition offers a formalized way for researchers of all stripes to compete in a meritocratic way – or, even better, you could set aside a space or two for teams that incorporate researchers from low- and middle-income countries – and that is another form of capacity building.”

As Williams of eGUIDE underscored, “we can’t do machine learning without high-quality data, and this requires that organizations can effectively collect, organize, store, and work with data. Data can transform the electricity sector, but capacity building is crucial.”