Reference Energy Disaggregation Data Set (REDD)

REDD contains power measurements from 6 real US homes collected over 3-19 days. The dataset provides two synchronized streams:

• Aggregate data: Whole-home power sampled at 1 Hz (smart meter equivalent) and 15 kHz (high-frequency).
• Circuit-level data: Individual appliance circuits sampled at 0.5 Hz, providing ground truth labels for up to 20 circuits per home.
• Mains voltage: Recorded to enable reactive power calculations.

REDD established the de facto benchmark for low-frequency NILM evaluation. The standard protocol dictates:

• Training on homes 1, 3, 4, and 5.
• Testing exclusively on homes 2 and 6 to assess generalization to unseen houses.
• Reporting F1-score, Mean Absolute Error (MAE), and Total Energy Correctly Assigned (TECA) for appliances like refrigerators, dishwashers, and microwaves.

REDD provides labeled ground truth for key household appliances with distinct consumption patterns:

• Continuous variable devices: Refrigerators and HVAC systems with cycling compressors.
• Multi-state appliances: Dishwashers and washing machines with sequential operational phases.
• On/off loads: Microwaves, disposals, and lighting circuits.
• Always-on loads: Electronics and standby consumption aggregated on specific circuits.

Researchers must account for REDD's documented imperfections:

• Missing data gaps: Homes 3 and 5 have significant periods of dropped aggregate readings.
• Unlabeled circuits: Several circuits remain unidentified or contain heterogeneous device groups.
• Short duration: 3-19 day collection windows limit the capture of seasonal or weekly behavioral patterns.
• Downsampling artifacts: The 1 Hz aggregate is derived from 15 kHz data, introducing slight timing offsets.

REDD catalyzed the shift from high-frequency transient analysis to low-frequency smart meter disaggregation. Key contributions include:

• Enabling the first reproducible comparisons between Factorial Hidden Markov Models (FHMM) and early deep learning approaches.
• Serving as the primary benchmark for Sequence-to-Sequence (Seq2Seq) and Denoising Autoencoder architectures.
• Establishing the precedent that public datasets are essential for advancing algorithmic rigor in energy disaggregation.

The process of identifying and isolating unique electrical characteristics that distinguish one appliance type from another in an aggregate power signal. Signatures fall into two categories:

• Steady-state signatures: Active/reactive power draw, harmonic content, and V-I trajectory shapes during stable operation
• Transient signatures: Inrush current spikes, startup duration, and spectral noise patterns during state transitions REDD's circuit-level measurements enable researchers to extract and validate these signatures against known appliance labels.

A generative statistical model representing an aggregate load as the sum of multiple independent hidden Markov chains, each modeling the state transitions of a single appliance. In the FHMM framework:

• Each appliance has discrete hidden states (e.g., OFF, ON, standby)
• The observed aggregate power is the sum of emissions from all active states
• Inference uses variants of the Viterbi algorithm or Gibbs sampling FHMMs were among the earliest algorithms benchmarked on REDD and remain a baseline for evaluating deep learning approaches.

A deep learning architecture that maps a sequence of aggregate power readings directly to a sequence of appliance-specific power consumption values. Unlike event-based methods, Seq2Seq models:

• Process entire windows of mains data (e.g., 512 samples) in one forward pass
• Use encoder-decoder structures with LSTM or Transformer backbones
• Capture long-range temporal dependencies without explicit state transition modeling REDD's time-series format makes it directly compatible with Seq2Seq training pipelines, and many state-of-the-art results on the dataset use this architecture.

Quantitative measures used to evaluate the performance of load monitoring algorithms on datasets like REDD. Key metrics include:

• F1-score: Harmonic mean of precision and recall for appliance state detection (ON/OFF classification)
• Mean Absolute Error (MAE): Average absolute difference between predicted and actual appliance power consumption in watts
• Total Energy Correctly Assigned (TECA): Percentage of total energy consumption correctly attributed to the right appliance
• Normalized Disaggregation Error (NDE): Ratio of disaggregation error to the error of always predicting the mean REDD's circuit-level ground truth enables precise calculation of all these metrics.

The algorithmic identification of significant state changes in an aggregate power signal—such as an appliance turning on or off—which serves as the primary trigger for many disaggregation pipelines. Event detectors typically:

• Monitor the derivative of the power signal for step changes exceeding a threshold
• Use matched filters or CUSUM (Cumulative Sum) algorithms for robust detection
• Classify events by magnitude, duration, and transient shape REDD's 1-second sampling rate at the aggregate level provides sufficient temporal resolution for event-based disaggregation methods to detect appliance state transitions.