Synthetic Aggregate Data Generation

The core mechanism involves linear superposition of individual appliance signatures to form a realistic aggregate signal. A synthetic aggregate $P_{agg}(t)$ is generated as:

$$P_{agg}(t) = \sum_{i=1}^{N} P_i(t) + \epsilon(t)$$

• $P_i(t)$: The power draw of appliance $i$ at time $t$, drawn from a library of real or simulated load profiles
• $\epsilon(t)$: Gaussian noise added to simulate sensor imprecision and unmodeled base loads
• Combinatorial explosion: With just 10 appliances each having 2 states, over 1,024 unique aggregate combinations exist, enabling massive dataset expansion from limited real recordings

Synthetic data must reflect human behavioral patterns and appliance usage rhythms to be useful for training. Key scheduling mechanisms include:

• Markov chain state transitions: Model the probability of an appliance turning on given the current time of day and previous state
• Seasonal and diurnal patterns: Heating loads peak in winter mornings; lighting loads dominate evenings
• Occupancy-driven activation: Simulated occupancy models trigger appliance usage based on the number of active residents
• Duration modeling: Appliances like dishwashers and washing machines follow characteristic runtime distributions rather than arbitrary on/off toggling

Without realistic scheduling, models trained on synthetic data fail to generalize to real-world consumption patterns.

Domain randomization deliberately varies parameters during synthesis to force NILM models to learn invariant features rather than memorizing specific signatures:

• Power level perturbation: Vary the steady-state power draw of an appliance by ±15% to simulate manufacturing tolerances and voltage fluctuations
• Transient shape warping: Apply random scaling and time-stretching to the high-frequency turn-on current spike
• Background load injection: Add continuously varying loads from unknown devices to prevent the model from relying on a clean residual signal
• Sampling rate augmentation: Train on mixtures of 1Hz, 1/6Hz, and 1/30Hz data to build sampling-rate-agnostic feature extractors

Models trained with domain randomization consistently outperform those trained on pristine synthetic data when deployed in real homes.

Beyond rule-based mixing, GAN-based synthesis learns to generate aggregate signals that are statistically indistinguishable from real measurements:

• Generator network: Takes random noise and conditioning labels (e.g., 'refrigerator + television + lighting') and outputs a synthetic aggregate waveform
• Discriminator network: Trained to distinguish real aggregate readings from generated ones, pushing the generator toward higher fidelity
• Conditional GAN architectures: Enable targeted generation of specific appliance combinations, useful for oversampling rare operational states
• Wasserstein loss: Used instead of standard GAN loss to stabilize training and avoid mode collapse when modeling the complex, multi-modal distribution of household power consumption

GAN-generated aggregates capture subtle correlations and noise characteristics that rule-based methods miss.

Synthetic aggregate generation enables privacy-compliant data sharing between utilities and researchers without exposing sensitive household information:

• No real occupancy data: Synthetic schedules are statistically representative but not tied to any actual household's behavior patterns
• Differential privacy guarantees: Noise can be calibrated to provide formal $\epsilon$-differential privacy bounds on the synthetic output
• Federated synthesis: Individual households generate synthetic data locally; only the synthetic aggregates are shared centrally for model training
• Regulatory compliance: Synthetic data falls outside GDPR and CCPA definitions of personal data when properly generated, enabling cross-border research collaboration

This approach addresses the fundamental tension between data utility and privacy that has historically limited NILM research.

The utility of synthetic aggregate data is measured by how well models trained on it perform on real-world disaggregation tasks:

• Train-on-synthetic, test-on-real (TSTR): The gold standard evaluation—train a NILM model exclusively on synthetic data and measure its F1-score on real household recordings
• Distribution similarity: Compare the synthetic and real aggregate distributions using Maximum Mean Discrepancy (MMD) or Wasserstein distance in feature space
• Event recall: Measure whether synthetic data preserves the frequency and characteristics of appliance state transitions
• Energy attribution accuracy: Verify that the total energy assigned to each appliance in the synthetic aggregate matches realistic consumption proportions

A low TSTR gap indicates high-fidelity synthesis; a large gap reveals that the synthetic data fails to capture critical real-world complexity.

A dual-network framework where a generator creates synthetic aggregate and appliance load signatures while a discriminator evaluates their statistical realism. This adversarial process enables the model to learn complex consumption distributions without requiring exhaustive real-world labeled data.

• Generator produces realistic active/reactive power sequences
• Discriminator distinguishes synthetic from real aggregate signals
• Enables generation of rare appliance combinations and edge cases
• Mitigates class imbalance in training datasets

The foundational process of isolating unique electrical characteristics—transient spikes, steady-state harmonics, and reactive power profiles—that distinguish one appliance from another. Synthetic data generation relies on accurate signature models to compose realistic aggregate signals.

• Transient signatures: inrush current patterns during startup
• Steady-state signatures: continuous harmonic distortion profiles
• Multi-state signatures: sequences for appliances with complex cycles (washer, dishwasher)
• Used as building blocks for bottom-up aggregate synthesis

A generative statistical model representing aggregate load as the sum of multiple independent hidden Markov chains, each modeling the state transitions of a single appliance. FHMMs serve as both a disaggregation technique and a synthetic data engine.

• Each appliance modeled as a discrete-state Markov process
• Aggregate signal = sum of individual state-dependent power draws
• Synthetic generation: sample state sequences, sum power outputs
• Captures realistic temporal dependencies and coincident operation

A widely-cited public benchmark containing low-frequency aggregate and circuit-level power measurements from multiple real homes. REDD serves as the ground truth foundation for training and validating synthetic data generation models.

• 6 homes with mains and individual circuit monitoring
• Sampling rates: 1 Hz (aggregate) and 15 kHz (high-frequency)
• Used to train generators that produce statistically similar synthetic data
• Standard benchmark for evaluating synthetic data fidelity

The algorithmic identification of significant state changes—on/off transitions, mode shifts—in an aggregate power signal. Synthetic data generators must accurately model the timing, magnitude, and shape of these events to produce realistic training sequences.

• Edge detection on power step changes (delta-P detection)
• Transient shape matching for appliance classification
• Synthetic events must replicate realistic inter-arrival times
• Critical for training event-based disaggregation pipelines

A methodology applying knowledge from a source domain to improve model performance in a target domain with limited labeled data. Synthetic aggregate data bridges the domain gap by providing diverse, labeled training examples that generalize across households.

• Pre-train on large synthetic datasets with known appliance labels
• Fine-tune on limited real-world data from target building
• Synthetic data introduces appliance diversity unseen in source
• Reduces need for expensive per-home circuit-level monitoring