Nairobi solar grids learn to forecast demand
Neighborhood batteries and better weather models are helping small operators plan for the evening power surge.
2 min read
The mini-grids were sized for lighting and phone charging. What arrived was refrigeration, welding, and a barber’s clippers, and the load curves stopped resembling anything in the design documents.
The forecasting problem
A mini-grid has no interconnection to fall back on. If demand exceeds what the array and battery can supply, the operator sheds load, and shedding load on a grid people paid to connect to is how an operator loses a village.
Forecasting a hundred-household grid is harder than forecasting a national one. There is no averaging: one welder starting up is a significant fraction of total demand, and demand is lumpy in a way that statistical methods built for aggregate systems handle badly.
On a small grid there is no crowd to hide behind. Every customer is a spike.
What actually improved accuracy
Not better weather models — those helped supply-side prediction and barely touched demand. The gains came from knowing what equipment had been connected.
Operators that maintain an appliance register, updated when a customer adds a machine, forecast substantially better than those relying on historical consumption alone. A welder that arrived last week has no history and a very predictable signature.
That turned a modelling problem into a customer-relationship problem: the register is only as good as the operator’s ability to get people to declare a new machine, which mostly comes down to whether declaring it costs them anything.
The operators with the best registers made declaration free and connection of an undeclared machine a tariff issue.
Productive use changes the business case
The grids that became financially viable are the ones where daytime commercial load grew. A grid serving only evening lighting has a terrible utilisation profile against a solar array that generates at midday.
Some operators now actively recruit daytime load — a mill, a cold store, a workshop — occasionally financing the equipment themselves, because a customer who uses power at noon is worth more than one who uses it at eight.
What to watch
Watch appliance register completeness rather than forecast error, since the first drives the second. Watch the ratio of daytime to evening consumption. And watch whether operators keep financing productive equipment when the first few loans go bad.
