Predicting the Grid That Doesn't Exist Yet: AI for African Power Utilities
African utilities forecast load for a grid still being built, serving customers whose consumption changes the moment supply becomes reliable.
African utilities are asked to forecast load for a grid that is still being built, serving customers whose consumption changes the moment supply becomes reliable. Historical demand curves therefore understate future demand systematically, and planning built on them guarantees under-capacity.
Useful forecasting here blends metered history with settlement growth, satellite night-lights, and the connection pipeline, rather than treating last year as prophecy.
Start with losses, because losses pay for the rest
Non-technical loss in several markets runs above twenty percent. Consumption-pattern anomalies, a meter whose profile flattens implausibly, a transformer whose downstream billed volume diverges from its measured throughput, identify tampering and unmetered connections far faster than field sweeps.
The condition is enforcement capacity. A list of two thousand suspect connections is useless to a utility that can inspect forty a week. Model output has to be tuned to the size of the field team, ranked by recoverable value, and routed as work orders rather than reports.
A list of two thousand suspect connections is useless to a utility that can inspect forty a week.
Forecasting for a demand curve that is still forming
Where reliability improves, latent demand appears: households that ran nothing but lighting begin using refrigeration; a trading centre adds welding and milling. Models trained purely on metered history cannot anticipate this, and utilities that plan on them chronically under-build.
Better inputs are structural: new connection pipelines, settlement expansion visible in satellite imagery, night-light intensity change, and the mix of commercial licences issued in a district. These are leading indicators of load in a way last year's kilowatt-hours are not.
The distributed side has better data than the grid
Pay-as-you-go solar operators hold high-frequency payment and usage telemetry that predicts default and hardware failure weeks ahead, and mini-grid data rich enough to size the next site properly. This is the most data-mature part of African energy, and the least connected to national planning.
The strategic opportunity is joining them: mini-grid consumption patterns are the best available evidence of what a newly connected community will actually demand from the main grid when it arrives.
Sequence deliberately
Loss detection first, because it produces cash within a financial year and builds internal credibility. Then demand forecasting, once the metering data has been cleaned as a by-product of loss work. Then generation and storage optimisation, which requires both.
Utilities that begin with optimisation models, usually because a vendor arrived with one, end up optimising against unreliable demand data and cannot demonstrate a return.
What this means in practice
- Deploy loss detection first and cap the output at the volume your field teams can actually inspect, ranked by recoverable revenue.
- Add structural leading indicators, connection pipeline, settlement growth, night-lights, to demand forecasting rather than relying on metered history.
- Formalise data sharing with pay-as-you-go solar and mini-grid operators to learn what newly connected demand looks like.
- Defer generation and storage optimisation until metering data quality has been measured and stated.
Frequently asked questions
- What is this analysis about?
- African utilities forecast load for a grid still being built, serving customers whose consumption changes the moment supply becomes reliable.
- What is the core argument?
- African utilities are asked to forecast load for a grid that is still being built, serving customers whose consumption changes the moment supply becomes reliable. Historical demand curves therefore understate future demand systematically, and planning built on them guarantees under-capacity.
- Start with losses, because losses pay for the rest?
- Non-technical loss in several markets runs above twenty percent. Consumption-pattern anomalies, a meter whose profile flattens implausibly, a transformer whose downstream billed volume diverges from its measured throughput, identify tampering and unmetered connections far faster than field sweeps.
- What should our organisation do about it?
- Deploy loss detection first and cap the output at the volume your field teams can actually inspect, ranked by recoverable revenue. Add structural leading indicators, connection pipeline, settlement growth, night-lights, to demand forecasting rather than relying on metered history. Formalise data sharing with pay-as-you-go solar and mini-grid operators to learn what newly connected demand looks like. Defer generation and storage optimisation until metering data quality has been measured and stated.
- Who published this and can we discuss it with Kwish?
- Kwish Research Team at Kwish Technologies published this on March 2026. Kwish works on energy programmes from offices in Uganda, Kenya, Sweden and Canada, and you can reach the team at info@kwishtechnologies.com.
Keep exploring
Turn this analysis into delivery
The services, sectors and case studies connected to this article.
Services that deliver this
Sectors where this applies
Related insights
Why African Governments Need AI-Native Systems, Not Retrofits
The temptation when modernising government technology is to retrofit. That instinct is exactly what Africa cannot afford.
The Hidden AI Infrastructure Gap in East African Universities
Faculty want AI in the classroom. IT departments don't have the data plumbing to deliver it. Here's the gap nobody is funding.
Building Government-Grade AI in Uganda: 5 Lessons from the Field
Procurement compliance. Stakeholder management. Data sovereignty. Local language requirements. Long-term maintenance planning.
Want this applied to your organisation?
Send us the decision you are trying to make. We reply with a scoped quotation.
Most quotations answered within one business day.