PowerUp · Kampala Site, Uganda

NILM Field Disaggregation Results

Electric Pressure Cooker

PowerUp detects 94% of EPC operating minutes and underestimates total energy consumption by just 0.1%

F1
95%
Precision
97%
Recall
94%
Energy error
−0.1%
EPC Field Disaggregation: Predicted vs actual appliance consumption at Kampala site, Uganda (W)
EPC disaggregation chart, Kampala site

Source: PowerUp 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 EPC cooking sessions, including periods of reduced electricity demand between heating cycles. This field-data example is one cooking session; across all unseen test days with valid plug data, the model detected 94% of EPC minutes and underestimated total energy by just 0.1%.

What it means

The model's ability to analyze longer consumption patterns helps capture different stages of an EPC cooking session.

Reactive power also helps distinguish appliances with similar power demands: Reactive power is close to zero for the EPC and negative for the induction stove, allowing the model to differentiate between the two.

With moderate confidence: errors across individual cooking sessions can cancel out, so the low overall error does not guarantee each session is accurate.

Limitations: Smart plugs had data on only 22–26% of days; per-session accuracy is not yet confirmed, so this is not yet reliable enough for appliance-specific billing.

Refrigerator

PowerUp's model detects fridge activity with an F1 score of 98%, but precision declines in homes with two fridges

F1
98%
Precision
98%
Recall
99%
Energy error
+0.5%
Fridge Field Disaggregation: Predicted vs actual appliance consumption at Kampala site, Uganda (W)
Fridge disaggregation chart, Kampala site

Source: PowerUp 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

This chart is from a one-fridge home. Refrigerators draw 60–150 W, far less than the 1,000–1,500 W cooking appliances, so PowerUp trained a separate model for them. Across all unseen test days in single-fridge homes, the model reached an F1 of 98%, energy within 0.5% of actual, and power within ±10% for at least 80% of running time.

What it means

A separate model for lower-power appliances improves detection and energy estimation in single-fridge homes.

In the one two-fridge home tested, precision drops to 94%: the model sometimes reports fridge activity when both are off, as it can't yet tell the two fridges apart. Moderate confidence, given this comes from only one two-fridge home.

Limitations: single-fridge results come from two homes; the two-fridge result from only one. Both need testing in more homes before wider use.

Induction Stove

PowerUp's model identifies 99% of induction stove operating minutes, with estimated energy consumption 4% below actual use

F1
98%
Precision
97%
Recall
99%
Energy error
−4%
Stove Field Disaggregation: Predicted vs actual appliance consumption at Kampala site, Uganda (W)
Stove disaggregation chart, Kampala site

Source: PowerUp 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.
1. Reactive power is a type of electrical power that helps reveal how different appliances operate. Its unique patterns can help distinguish appliances with similar electricity consumption.

What we're seeing

This field-data example is one of many cooks; across all unseen test days, the model detected 99% of operating minutes and underestimated energy by about 4% in both households. When the stove operates, its reactive power drops below zero, a distinct pattern not seen in other appliances tested.

What it means

Reactive power1 helps the model distinguish the induction stove from the EPC, despite their similar power demands. The consistent 4% underestimate across both households suggests a downward bias rather than errors cancelling each other out. With moderate confidence, given the bias held across both households: using this for appliance-specific billing would undercharge customers and reduce operator revenue.

Limitations: Only two homes had a stove; simultaneous use with other cooking appliances is not yet tested.