Can one electricity meter tell you what appliances are running in a home?

A standard meter records total consumption and cannot distinguish between electricity used by a refrigerator, an induction stove or a maize mill. Currently, obtaining this level of detail requires a dedicated submeter costing $30–40 per appliance. Access to appliance-level consumption data could enable mini-grid developers to introduce tariffs based on how and when electricity is used, encouraging clean cooking and improving electricity utilization.

This work builds on the Lab’s October 2026 smart metering diagnostic, which found that metering can account for up to 15% of capex and 64% of opex per connection, and that developers want a basic meter at $20–40. A single appliance submeter can therefore cost as much as an entire household meter.

Non-intrusive load monitoring (NILM) explores whether a single household meter can provide this information. Using machine learning and AI algorithms, NILM analyses a household’s total electricity consumption to estimate when individual appliances are running and how much electricity they consume. In 2025, the Lab’s first NILM research with PowerPay tested this concept in the lab. In 2026, we tested whether the approach works in real households.

What we are testing

The primary objectives of the study are:

• Determine whether mini-grid operators can use a single, cost-effective “multi-load meter” to identify different loads
• Determine the viability of implementing type-of-use and time-of-use tariffs using such a device, and the impact of these tariffs on electricity demand

What we did

With funding from Transforming Energy Access (TEA), PowerPay (Kiambu, Kenya) and PowerUp (Kampala, Uganda) each tested their own model in real homes in the field. PowerPay uses a convolutional neural network (CNN) to identify individual appliance usage patterns within a household’s total electricity consumption, such as a fridge switching on and off, an electric pressure cooker holding pressure or a mill’s high-power burst. PowerUp uses a transformer-based NILM model that analyses one-hour windows of household electricity consumption data, with power and voltage measurements recorded every second.

We checked every prediction against actual appliance readings, using submeters at PowerPay sites and smart plugs at PowerUp sites, on test days the models had not seen in training. The appliances were electric pressure cookers (EPCs), induction stoves, refrigerators and one electric mill.

How we scored the models

We measured detection and energy accuracy separately. We assessed detection accuracy using the F1 score, which combines two questions: when the model says an appliance is on, how often is it right (precision), and of the minutes the appliance actually ran, how many did the model find (recall). A score of 100% is perfect. The F1 score does not measure how accurately the model estimates energy consumption, so we also report energy-estimation error, where a negative figure means the model undercounted.

Results

Both models were able to disaggregate loads and identify different appliances successfully. Detection accuracy was particularly high for cooking appliances and refrigerators. EPCs and induction stoves reached F1 scores of 95–98%. Refrigerators reached 98–99%. Energy errors ranged from a 7% underestimate to a 0.5% overestimate. The model performed less accurately for the mill: an F1 of 84%, 76% of operating minutes detected and total energy underestimated by 36%.

Explore the field results for PowerPay and PowerUp, including graphs comparing predicted and actual appliance consumption, key observations, and what the results mean for NILM applications in mini-grids.

What the evidence does not yet show

While the initial results are promising, several limitations affect how confidently the findings can be applied beyond the households tested.

• Appliance submeters recorded data on only 8–25% of days at PowerPay sites. Smart plugs at PowerUp sites covered 22–62% of days.
• The mill result rests on one mill and 15 reference days.
• Only two homes had an induction stove, and only one had two fridges. In that home, precision fell to 94%.
• The EPC and mill ran together on only one day at PowerPay sites, so we could not properly test simultaneous use.
• Neither model has been tested in a household it was not trained on.
• The current energy estimates are not accurate enough for appliance-specific or time-specific billing. More testing and data are needed before differentiated tariffs can be applied

Next steps

Together with PowerUp and PowerPay, we are continuing to test the models across additional households and appliances to improve their accuracy and strengthen the evidence base. A more detailed analysis of the findings will be published soon.

Funded by