PowerPay · Kiambu Site, Kenya

NILM Field Disaggregation Results

Electric Pressure Cooker

PowerPay accurately detects EPC cooking activity, identifying 98% of operating minutes and estimating energy within 6%

F1
97%
Precision
96%
Recall
98%
Energy error
−6%
EPC Field Disaggregation: Predicted vs actual appliance consumption at Kiambu site, Kenya (W)
EPC disaggregation chart, Kiambu site

Source: PowerPay NILM report, Oct 2026.
Note: This visualization shows one example day from the unseen field test data. The metric scores above (F1, Precision, Recall, Energy error) are aggregated across all unseen field test days with valid submeter data.

What we're seeing

The model follows two long cooking sessions, from about 10:47am to 11:15am and from 12:03pm to 12:20pm, when the EPC draws 700–850 W. Its on/off periods match the submeter almost exactly, though it misses some lower-power draw between heating cycles. This field-data morning is one example; across all unseen test days with valid submeter data, the model found 98% of EPC minutes and estimated energy within 6%.

What it means

The EPC's distinct, high-power pattern makes it easy to identify when cooking occurs.

With moderate-to-high confidence, NILM can support monitoring of EPC usage and how often households cook; EPC submeters had valid data on only 11–25% of days, so this rests on a limited sample.

Limitations: Not yet tested in new homes; estimates are not accurate enough for appliance-specific billing.

Refrigerator

PowerPay achieves a 99% F1 score for refrigerator detection, but underestimates energy consumption by 7%

F1
99%
Precision
99%
Recall
99%
Energy error
−7%
Fridge Field Disaggregation: Predicted vs actual appliance consumption at Kiambu site, Kenya (W)
Fridge disaggregation chart, Kiambu site

Source: PowerPay NILM report, Oct 2026. One mill, 15 submetered days.
Note: This visualization shows one example day from the unseen field test data. The metric scores above (F1, Precision, Recall, Energy error) are aggregated across all unseen field test days with valid submeter data.

What we're seeing

Overnight, the fridge switches on and off every few minutes at about 45–50 W, beneath a household load of 100–200 W. This field-data night is one example; across all unseen test days, the model found 99% of fridge minutes and estimated energy 7% below actual.

What it means

The refrigerator's distinct switching pattern supports accurate detection, but its relatively low power demand makes small estimation errors more significant.

With low-to-moderate confidence, NILM could support refrigerator-use monitoring: reference data was available on only 8% of days, so this result rests on little evidence.

Limitations: Small sample of valid reference days; the 7% energy gap is too large for appliance-specific billing; not yet tested on more refrigerators.

Mill

PowerPay detects 76% of mill operating minutes, resulting in a 36% underestimate of electricity consumption

F1
84%
Precision
94%
Recall
76%
Energy error
−36%
Mill Field Disaggregation: Predicted vs actual appliance consumption at Kiambu site, Kenya (W)
Mill disaggregation chart, Kiambu site

Source: PowerPay NILM report, Oct 2026. One mill, 15 submetered days.
Note: This visualization shows one example day from the unseen field test data. The metric scores above (F1, Precision, Recall, Energy error) are aggregated across all unseen field test days with valid submeter data.

What we're seeing

The mill runs once on this test afternoon, for about eight minutes from 15:00, peaking at 2,000–2,500 W. This field-data afternoon is one example; the model found this across all unseen test days, though it misses a short start at 14:55pm. While the model achieves 94% precision when identifying mill activity, it detects only 76% of actual operating minutes, resulting in a 36% underestimate of total mill energy consumption.

What it means

With low confidence, the model struggles to consistently identify short mill runs: with only one mill and 15 reference days, it had few examples to learn from and validate against.

Limitations: One mill, 15 reference days; the 36% energy gap is too large for productive-use billing; more mills and data are needed before drawing conclusions.