Harvesting the synergies of machine learning and survey data for electrification planning in Ethiopia

Background, challenges, and context

While Ethiopia is currently undertaking a National Electrification Program (NEP 2.0), research suggests that rural households, which comprise 80% of Ethiopia’s population, will have very low consumption patterns for quite some time. Effective planning tools are pivotal in defining the most suitable electrification strategy for different sub-national regions, comparing the extension of the national grid to the establishment of mini-grids or other off-grid systems.

Demand forecasting models do not yet account for potential productive use hubs, including agricultural demand in specific regions. The agricultural sector is the backbone of large parts of the rural African economy, responsible for a large share of self-employed and employed workers. Fostering electricity consumption for productive purposes, especially in the agricultural sector (and thereby improving revenue collection), will be key for policy makers to assure long-term sustainability of grid access policies.

The Ministry of Water and Energy (MoWE) and the Ministry of Agriculture (MoA) have repeatedly sought more research on electrification planning related to productive use, and particularly agriculture.

Research overview and objectives

This research project aimed to inform stakeholders in the agricultural sector on understanding, forecasting, and prioritising electricity demand for agricultural purposes. The research questions were:

• How can a utility develop effective electricity demand forecasts and stimulation techniques for productive uses in agriculture?

• How can machine learning techniques best be combined with classical on-the-ground surveys to yield informative demand forecasts regarding productive uses in agriculture that ultimately also inform electricity system expansion? Can such techniques be a cost-effective method for enhancing electricity demand forecasts?

• Which agricultural demand stimulation interventions related to irrigation and agro-processing can be derived from these demand forecasts?

Irrigation can increase the productivity of agricultural production, and agricultural processors can modernise their operations by replacing manual and diesel-fuelled machines with electric appliances. Further, reducing the use of diesel-powered irrigation pumps could reduce dependence on fossil fuels, help achieve Ethiopia's climate action plan, and yield cost savings among farmers.

The team aimed to develop approaches that would help identify potential high-demand regions where the grid roll-out should be concentrated. The project combined an innovative machine learning prototype simulation model with classical ‘on-the-ground’ surveys among households, enterprises, and communities.

Research methodology

The aim was to develop a scalable, cost-effective approach that uses relatively ‘expensive to collect’ information by carrying out surveys in a limited number of regions and then train machine algorithms to extrapolate this in-depth information.

The team collected real examples of where irrigation was taking place in Ethiopia's Amhara and Oromia regions as well as ground-truth information on the method of irrigation used. Areas with existing diesel-powered irrigation were identified by combining data from an agricultural survey with satellite-measured pollution, crop cover, and topography data. This approach leveraged an essential characteristic of diesel pumps; they emit pollutants that can be measured remotely by a satellite instrument. Furthermore, it is based on the hypothesis that diesel-irrigation activity is detectable by matching pollution patterns to irrigation seasonality.

The areas were clustered into 250m patches and labelled based on the prevalent irrigation method (diesel- or non-diesel). The team trained a machine learning model to distinguish between diesel-based and non-diesel-based irrigated areas. Satellite imagery was used to predict the likelihood of irrigation.

In addition, quantitative and qualitative data was collected from households, enterprises, communities, and regional representatives in the two aforementioned regions with particular potential for irrigation: an area east of Lake Tana in the Amhara region and central parts of the Oromia region, each covering around 3,000km2. This sought to ‘ground-truth’ the satellite analysis and understand directly the opportunities and barriers for agricultural and non-agricultural productive uses of electricity.

Research results, key messages, and recommendations

Satellite data and machine learning

• Predictive data analytics using publicly accessible satellite data can be used to identify locations with latent electricity demand for irrigation with 75% accuracy, potentially offering utilities a consistent and substantial source of revenue to supplement the low residential demand from rural customers.

• Though the study showed promising results, the limited ground-truth sample size collected, and the nature of the publicly available satellite data used, makes it difficult to generalise the findings.

• To enhance the use of satellite data and machine learning for data-driven electrification planning, governments and other private actors should increase efforts to make high-quality ground-truth data from surveys, energy systems, and other technologies accessible and available to researchers to overcome the barrier of collecting large ground-truth samples, which is highly resource-intensive and logistically challenging.

• Higher resolution imagery could result in more extensive and higher quality ground-truth samples.

• Researchers and policy makers should collaborate to learn how to make satellite data and predictive data analytics outputs interpretable and actionable.

Productive use of grid electricity in relation to small-scale irrigation

• Almost all non-connected enterprises in the rural Ethiopian sample considered the lack of electricity as a severe or very severe obstacle to their business operations.

• According to farmers in the study regions, water scarcity currently seems to be the larger obstacle to irrigation in comparison to the lack of electricity, along with the high per unit fuel cost of diesel motorised water pumping.

• Few farmers adopt electricity for irrigation, not only because of switching costs from diesel to grid-electricity motorised pumps, but primarily because grid electricity is unavailable at their fields.

• Rural grid coverage would need to be expanded considerably to increase the use of electricity for irrigation.

• The high portability of diesel pumps makes them the technology of choice among farmers in the study sample regions and beyond.

• Decentralised solutions such as solar-powered pumps may prove to be a more cost-effective and flexible solution in some settings. The daily and seasonal electric load shapes of irrigation must also be considered.

• Despite the presence of solar power in enterprises where grid connection is low or unavailable, grid connections seem to be critical to ensure the introduction of higher-power appliances.

• The range of enterprises and scope of new productive electricity uses in electrified communities remains limited. Integrated rural development strategies that improve market access more generally are required.

• Existing research supports the overall picture emerging from this study, of mixed evidence on the opportunities of productive use from grid electrification in rural areas, especially in the agricultural sector.

The team worked with electricity system planners and analysts from the Electrification Directorate in MoWE to incorporate the findings into electricity planning processes.

Local partner

Policy Studies Institute (PSI)