Universal NILM (UNILM): Unsupervised Energy Disaggregation

• Universal NILM (UNILM) is an unsupervised framework that disaggregates aggregate household power into individual appliance signatures without requiring labeled data.
• It employs dual architectures—a probabilistic filter pipeline with knapsack optimization and an LLM-based prompting method—to achieve accurate, transparent event disaggregation.
• The framework demonstrates scalable, region-agnostic performance with competitive metrics, achieving over 93% accuracy in appliance energy allocation on benchmark datasets.

Universal NILM (UNILM) constitutes a class of fully unsupervised, region-agnostic frameworks for non-intrusive load monitoring, targeting the disaggregation of aggregate household electricity consumption into individual appliance signatures and energy usage without the need for prior labeled data or hand-tuned regional models. UNILM approaches have advanced along two primary axes: probabilistic signal processing pipelines with online appliance discovery (Rodriguez-Silva et al., 2019), and prompting-based, LLM architectures for training-free disaggregation with native explainability (Xue et al., 9 May 2025).

1. Problem Definition and Objectives

Universal NILM (UNILM) is formulated to operate without supervised training, explicit prior appliance models, or region-specific feature engineering. The fundamental challenge addressed is automatic, unsupervised detection, tracking, and semantic labeling of appliances from raw, aggregate power time series under highly variable regional appliance mixes and usage patterns. The objectives are:

• Fully unsupervised appliance discovery and tracking from aggregate signals.
• Online operation suitable for embedded or edge hardware.
• Region-agnostic transferability via semantic partition mapping.
• Quantitative disaggregation at high aggregate energy accuracy.
• Transparency and explainability in appliance state and energy attribution.

2. Methodological Frameworks

There exist two prominent UNILM realizations:

2.1 Filter-Pipeline plus Probabilistic Knapsack UNILM

The architecture (Rodriguez-Silva et al., 2019) comprises three principal stages:

1. Multi-Stage Filter Pipeline: Raw aggregate power $z_t$ is processed through (i) median filtering, (ii) bilateral filtering, (iii) anisotropic diffusion (Perona–Malik), (iv) one-dimensional domain transform, and (v) edge sharpening. The output $y_t = f(z_t)$ retains event edges and steady-states while suppressing noise.
2. Unsupervised Appliance Modeling and Event Disaggregation: ON/OFF events are detected via thresholds on $\Delta y = y_t - y_{t-1}$ ($s$=60 W). Disaggregation leverages a probabilistic multiple-choice knapsack problem (MCKP). Each discovered appliance $\alpha_i$ is modeled by four Gaussians:

$$\alpha_i = \{ \mathcal{N}_{P_0}, \mathcal{N}_{P_1}, \mathcal{N}_{D_0}, \mathcal{N}_{D_1}, y_i[1...T] \}$$

where $\mathcal{N}_{P_1}$ models ON-state power, $\mathcal{N}_{P_0}$ OFF-state/standby, and $\mathcal{N}_{D_{1,0}}$ ON/OFF durations.

Appliance-ON/OFF assignments solve:

$$\text{maximize} \ \sum_{i=1}^M \sum_{j \in N_i} p_{ij} x_{ij} \ \text{subject to} \ \sum_{i=1}^M \sum_{j \in N_i} w_{ij} x_{ij} \leq |\Delta y| ; \ \sum_{j \in N_i} x_{ij} = 1 \ \forall i; \ x_{ij} \in \{0,1\}$$

with $y_t = f(z_t)$0 (candidate power step) and $y_t = f(z_t)$1 proportional to the pdf value $y_t = f(z_t)$2.

1. Semantic Labeling via Partition Map ($y_t = f(z_t)$3): Discovered appliances are mapped to semantic labels based on their region, event duration $y_t = f(z_t)$4, and magnitude $y_t = f(z_t)$5, using a prebuilt 3D lookup $y_t = f(z_t)$6 to assign classes like “fridge” or “dryer”.

2.2 LLM-Prompting-Based UNILM

The prompting-based variant (Xue et al., 9 May 2025) eliminates specialized signal models in favor of a LLM guided by engineered prompts:

• Preprocessing: Mains readings $y_t = f(z_t)$7 are resampled (6 s), missing data interpolated, and per-appliance states $y_t = f(z_t)$8 thresholded.
• Prompt Construction: The prompt includes system/task definitions, appliance statistics (standby, ON-range, cycle), one-shot in-context examples ($y_t = f(z_t)$9 or $\Delta y = y_t - y_{t-1}$0) pairing aggregate input to true state labels, timestamps, user queries, and (optionally) context from prior windows.
• LLM Inference: The model infers $\Delta y = y_t - y_{t-1}$1 in structured JSON, and can append natural-language explanations for each ON/OFF decision.
• Post-processing: Outputs are rectified for length and evaluated against ground truth.

3. Mathematical and Algorithmic Elements

Probabilistic Knapsack (MCKP)

Each ON/OFF event triggers a discrete MCKP optimization over candidate appliance step values ($\Delta y = y_t - y_{t-1}$2 from $\Delta y = y_t - y_{t-1}$3 bands of $\Delta y = y_t - y_{t-1}$4). The modeled probability that appliance $\Delta y = y_t - y_{t-1}$5 generated $\Delta y = y_t - y_{t-1}$6 is

$\Delta y = y_t - y_{t-1}$7

Assignment and cluster growth are governed by profit ($\Delta y = y_t - y_{t-1}$8 for confident attribution) and the Mahalanobis distance ($\Delta y = y_t - y_{t-1}$9 for forming a new appliance cluster).

LLM Prompt Templates and Inference

Prompts specify: task role, appliance knowledge, time-series input/output examples, and request status lists. Precision, recall, and F1 are evaluated per appliance $s$0:

$s$1

Average F1 summarizes overall model performance.

4. Empirical Results and Benchmarking

Filter-Pipeline UNILM (Rodriguez-Silva et al., 2019)

Experiments on RAE House-1, Block-1 (1 Hz, 9 days) demonstrate:

• Three appliances discovered (dryer, fridge, furnace).
• Aggregate disaggregation achieves $s$2 accuracy for energy allocation:
  • Clothes Dryer: $s$3
  • Fridge: $s$4
  • Furnace: $s$5
• Total run time for a day: $s$614.9 min (2.3 GHz Core i5).
• No prior knowledge or labelled data required; region transfer only alters $s$7.

LLM-Prompting UNILM (Xue et al., 9 May 2025)

Evaluation on REDD House-1 (low-frequency mains, unseen test):

• LLM variant (DeepSeek-V3) matches state-of-the-art sequence-to-point (Seq2Point) models (average F1 = 0.676) without any parameter fine-tuning.
• Particularly strong on event-driven appliances (microwave, washer).
• LLM output natively delivers stepwise explanations, increasing transparency compared to black-box deep learning.

Model	Microwave	Fridge	Washing Machine	Dishwasher	Avg F1
Seq2Seq	0.498	0.753	0.670	0.556	0.619
Seq2Point	0.543	0.778	0.757	0.628	0.676
GPT-4.1-mini	0.634	0.285	0.704	0.466	0.602
DeepSeek-V3	0.621	0.534	0.740	0.568	0.676

5. Scalability, Transferability, and Regional Adaptation

UNILM frameworks are explicitly constructed for scalable and region-agnostic operation:

• Scalability: Linear complexity in time ($s$8) and near-linear in appliances ($s$9) and candidate steps per event. Knapsack heuristics enable real-time execution ($\alpha_i$0 Hz) on embedded platforms (Rodriguez-Silva et al., 2019).
• Zero-shot Adaptation: LLM-based UNILM requires no retraining; adaptation across regions is handled via prompt reconfiguration (statistics, context). Signal-processing UNILM only mandates regional updates to the partition map $\alpha_i$1 (Xue et al., 9 May 2025).
• Online Discovery: New appliances are allocated cluster models as dictated by clustering distance and knapsack solution confidence. Gaussian parameters are updated via exponential averaging.

6. Explainability and Transparency

A distinctive feature of LLM-based UNILM architectures is the capability to emit structured, human-readable explanations for the current ON/OFF status at each time step:

• Explanations can cite lack of characteristic power surges, absence of known cyclic patterns, or direct correspondence to appliance cycle statistics: $\alpha_i$2 This addresses longstanding critique of explainability in deep NILM models (Xue et al., 9 May 2025).

7. Impact and Extensions

UNILM establishes a practically deployable, training-free NILM pipeline requiring no labelled datasets, suitable for global rollouts and cross-country deployment. The generalization of appliance attribution performance is competitive with or exceeds specialized deep sequence models on publicly available benchmarks (Xue et al., 9 May 2025). By decoupling detection/disaggregation from semantic appliance identities, and leveraging region-specific label lookup, UNILM’s flexibility exceeds conventional supervised NILM approaches. A plausible implication is that UNILM methodologies—particularly LLM-based prompting with knowledge injection and in-context example construction—may extend across broader event disaggregation and unsupervised segmentation domains beyond residential energy analytics.