Low-Frequency NILM

The defining characteristic is a sampling interval of 1 second or greater, often aggregating to 1-minute or 15-minute intervals. This low data rate captures only steady-state active power changes, making the problem computationally lighter but information-poor compared to high-frequency methods. The data is typically a single scalar value (Watts) per time step, losing all harmonic and transient information.

Without high-resolution waveforms, low-frequency NILM fundamentally relies on event-based detection. Algorithms scan the aggregate power signal for significant step changes in active power (ΔP). A positive step indicates an appliance turning on; a negative step indicates an off event. The core challenge is matching these magnitude changes to a database of known appliance signatures, a process complicated by overlapping events and small power draws.

The primary feature for identification is the steady-state power draw in Watts. An appliance is modeled as a finite-state machine where each operational state has a characteristic power consumption value. Disaggregation becomes a combinatorial optimization problem: finding the subset of known appliance states whose power values sum to the observed aggregate reading. This approach struggles with continuously variable devices like dimmed lights or inverter-driven heat pumps.

To disambiguate appliances with similar power draws, low-frequency algorithms heavily weight temporal features:

• On-duration histograms: How long a device typically runs.
• Duty cycles: The ratio of on-time to total cycle time for cyclic loads like refrigerators.
• Time-of-day priors: Bayesian models incorporate the probability that a device operates at a specific hour, distinguishing a 2kW oven from a 2kW space heater based on usage patterns.

Modern approaches replace FHMMs with neural networks that learn complex temporal dependencies directly from data:

• Sequence-to-Sequence (Seq2Seq): An encoder LSTM compresses a window of aggregate power into a context vector; a decoder LSTM generates the target appliance's power trace.
• Temporal Convolutional Networks (TCN): Use dilated convolutions to capture long-range dependencies with a larger receptive field than RNNs.
• Transformers: Apply self-attention to model relationships between any two time steps, effectively handling appliances with long off-periods.

Utility companies integrate low-frequency NILM into customer portals to provide itemized billing and personalized efficiency tips. By analyzing 15-minute or hourly smart meter readings, algorithms identify major loads like HVAC systems, electric water heaters, and pool pumps. This behavioral feedback has been shown to reduce household consumption by 2-5% through increased awareness alone.

• Example: Bidgely and Oracle Utilities deploy NILM engines that parse AMI data to show customers exactly how much their air conditioner cost last month.
• Key metric: Disaggregation at 15-minute resolution achieves >90% accuracy for large resistive loads.

Low-frequency NILM enables utilities to identify specific flexible loads behind the meter without customer surveys. Knowing which homes have electric vehicles or heat pumps allows for precise demand response enrollment. During peak events, operators can send signals to curtail identified loads while verifying compliance through continued disaggregation of the aggregate signal.

• Mechanism: Step-change detection in 1-second to 1-minute data reveals EV charging onset, triggering automated load shifting.
• Benefit: Avoids blanket load shedding, preserving comfort for non-participating appliances.

Energy efficiency programs traditionally require expensive on-site inspections to verify measure installation. Low-frequency NILM provides remote verification by detecting changes in appliance operation patterns pre- and post-retrofit. A new high-efficiency refrigerator exhibits a distinct duty cycle signature compared to the replaced unit.

• Application: Pay-for-performance programs use disaggregation to confirm savings without entering the home.
• Scale: One model can verify thousands of retrofits simultaneously using existing AMI infrastructure.

By tracking the duty cycle and power draw trajectory of compressors and air handlers over weeks, low-frequency NILM detects degradation before catastrophic failure. A gradual increase in startup power or longer run-times to achieve the same temperature setpoint signals refrigerant leakage or mechanical wear.

• Signal analyzed: Trend analysis of step-change magnitude and on-duration from hourly smart meter data.
• Outcome: Proactive service dispatch reduces emergency repair costs and prevents comfort complaints.

Multi-family buildings and commercial properties use low-frequency NILM to allocate utility costs fairly without installing per-unit meters. The aggregate building load is disaggregated to estimate individual tenant consumption based on identified appliance usage patterns and occupancy schedules.

• Regulatory context: Complies with tenant submetering laws in jurisdictions that permit algorithmic allocation.
• Hardware requirement: A single smart meter at the building service entrance replaces dozens of downstream sensors.