Innovation Insight: Machine Learning to Predict Mini-Grid Consumption
One of the fundamental challenges to the profitability of mini-grids is the difficulty of predicting customer demand, and the high costs associated with building a grid with either too much or too little generating capacity. For example, if a 30 kW mini-grid achieves a 12.5% internal rate of return (IRR) when capacity perfectly matches demand, then when oversized by 50% IRR drops to 9.2%, and when oversized by 100% IRR drops to just 6.1%. Also, when grids are undersized and built with too little generating capacity to meet customer demand, developers either miss out on potential revenue, or incur high costs by retroactively expanding the grid or using expensive diesel to supplement renewable electricity generation. One of the Lab’s developers recently described accurately forecasting demand at new sites as “the single biggest challenge developers face.”
What makes it so difficult to accurately predict consumption at a site? Electricity demand is influenced by numerous characteristics of a community; it is challenging to isolate the relative importance of each of these and how they interact with each other to impact demand. A site’s local economy affects its residents’ need for electricity. For example, a village on Lake Victoria may have a high demand for charging fishing lights that allow people to fish at night. Demographic and socioeconomic characteristics also affect customers’ willingness and ability to pay. Older customers may see less need for electricity, since they’ve lived so long without it, and may have less disposable income to pay for it.
However, as ever more consumption and payment data is combined with increasingly granular information on site characteristics, machine learning techniques should enable past experience to predict the future. With over 550 million data points on customer behavior across 62 sites in 4 countries, the Lab is uniquely placed to help mini-grid developers use historic patterns to generate accurate estimates of energy demand at new sites.
In 2019, the Lab initiated this project by working with DataKind, a nonprofit which conducts data science projects with mission-driven organizations, to produce a basic model to predict mini-grid customers’ consumption from survey responses. To learn which customer characteristics most inform consumption, and how, we combined almost two years of mini-grid consumption and payment data from 31 sites across East Africa with customer survey responses. The goal was to produce a model which would allow developers to forecast electricity demand using streamlined customer surveys or census data.
The resulting analysis shows each customer’s survey responses are not predictive of their individual consumption. The Lab’s best performing model, using random forest regression, predicted each customer’s demand with a 65% error rate – i.e. if a customer’s true consumption was 10 kilowatt hours (kWh)/month, the model might predict their consumption to be 16.5 kWh/month or 3.5 kWh/month.