Lightweight and Scalable Transfer Learning Framework for Load Disaggregation

L. E. Garcia-Marrero; G. Petrone; E. Monmasson

cs.LGMar 5, 2026

Lightweight and Scalable Transfer Learning Framework for Load Disaggregation

Authors: L. E. Garcia-Marrero, G. Petrone, E. Monmasson

Organizations: Systèmes et Applications des Technologies de l’Information et de l’Energie laboratory, Cergy-Pontoise · Dipartimento di Ingegneria dell’Informazione ed Elettrica e Matematica Applicata/DIEM, University of Salerno, 84084 Fisciano

Abstract

Non-Intrusive Load Monitoring (NILM) aims to estimate appliance-level consumption from aggregate electrical signals recorded at a single measurement point. In recent years, the field has increasingly adopted deep learning approaches; however, cross-domain generalization remains a persistent challenge due to variations in appliance characteristics, usage patterns, and background loads across homes. Transfer learning provides a practical paradigm to adapt models with limited target data. However, existing methods often assume a fixed appliance set, lack flexibility for evolving real-world deployments, remain unsuitable for edge devices, or scale poorly for real-time operation. This paper proposes RefQuery, a scalable multi-appliance, multi-task NILM framework that conditions disaggregation on compact appliance fingerprints, allowing one shared model to serve many appliances without a fixed output set. RefQuery keeps a pretrained disaggregation network fully frozen and adapts to a target home by learning only a per-appliance embedding during a lightweight backpropagation stage. Experiments on three public datasets demonstrate that RefQuery delivers a strong accuracy-efficiency trade-off against single-appliance and multi-appliance baselines, including modern Transformer-based methods. These results support RefQuery as a practical path toward scalable, real-time NILM on resource-constrained edge devices.

Figures & tables

Fig. 1: Three-stage RefQuery framework. Stage I: the model is trained on multiple source buildings using several appliances per building with joint power (MSE) and state (BCE) supervision; Stage II: in the target domain, the model is frozen and a target-specific reference embedding is learned per appliance using joint power (MSE) and state (BCE) supervision; Stage III: target mains is disaggregated into per-appliance state and power using the learned embeddings.

Fig. 2: Compact 1D CNN feature extractor for embedding generation.

Fig. 3: Multitask head architecture. Given the reference and query embeddings $\mathbf{e}_{r},\mathbf{e}_{q}\in\mathbb{R}^{E}$ , the interaction module constructs a joint representation by concatenating $[\mathbf{e}_{r},\;\mathbf{e}_{q},\;(\mathbf{e}_{q}-\mathbf{e}_{r})^{2},\;\mathbf{e}_{q}\odot\mathbf{e}_{r}]$ , where $\odot$ denotes element-wise multiplication. The resulting $4E$ -dimensional vector is processed by two shared fully connected layers with ReLU activations and width $H$ to produce a shared representation. Two task-specific branches are then applied: a classification head (dense + sigmoid) predicts the state probability $\hat{y}_{\text{state}}\in\mathbb{R}$ , and a regression head (dense + ReLU) estimates a non-negative magnitude $\Delta z\in\mathbb{R}$ . The final standardized power estimate is computed through a gated transformation $\hat{y}_{z}=\Delta z\,\hat{y}_{\text{state}}+b_{\text{off}}$ , where $b_{\text{off}}=-\mu/\sigma$ anchors the OFF operating point in $z$ -space.

Metric	Method	UK-DALE (% change)
Metric	Method	F	DW	MW	WM	K	Ave
$\Delta$ MAE	SA-S2P	41.2	165.1	6.7	65.9	118.3	63.5
	MA-CNN	52.3	132.5	39.0	96.0	81.8	70.5
	MA-TRF	7.8	54.6	3.9	30.4	60.9	23.3
	RefQuery	22.1	18.1	-1.3	15.6	14.9	16.1
$\Delta$ F1	SA-S2P	-15.9	-40.2	87.5	-26.1	-40.2	-23.6

TABLE V: Relative changes on UK-DALE (Limited Data vs. Normal) for House 1, in %. Positive

\Delta

MAE indicates worse error; negative

\Delta

F1 indicates worse accuracy. Best (per appliance/column) is in bold.

Metric	Method	REDD House 1 (% change)
Metric	Method	F	DW	MW	WM	Ave
$\Delta$ MAE	SA-S2P	30.7	47.3	12.5	8.3	21.3
	MA-CNN	26.5	103.8	69.3	-8.1	31.2
	MA-TRF	15.1	6.8	19.6	19.3	16.2
	RefQuery	16.2	12.4	38.4	8.5	18.9
$\Delta$ F1	SA-S2P	-7.4	-12.3	-17.8	-50.8	-19.9

TABLE VII: Relative changes on REDD (Limited Data vs. Normal) for House 1, in %. Positive

\Delta

MAE indicates worse error; negative

\Delta

F1 indicates worse accuracy. Best (per appliance/column) is in bold.

Method	Storage (MB) $\downarrow$	Adaptation (s) $\downarrow$	Inference (s) $\downarrow$
SA-S2P	334.50	5400.3	391.07
MA-CNN	2.38	4950.5	115.11
MA-TRF	11.81	4429.5	248.86
RefQuery	1.18	63.2	42.41

TABLE VIII: Computational comparison of the methods for single-appliance adaptation using 7 days of data, followed by one-month offline inference, evaluated under a CPU-only constrained execution regime (process affinity limited to two CPU cores and thread-level parallelism limited to two threads).

Method	$\Delta W$ (KB) $\downarrow$	Incremental?	Inference scaling (MFLOPs) $\downarrow$
SA-S2P	$3.4\times 10^{5}$	Yes a	$0.000+K\cdot 101.289$
MA-CNN	$\approx 1$	No c	$13.295+K\cdot 0.000129$
MA-TRF	$4.6\times 10^{3}$	Yes b	$889.560+K\cdot 1142.726$
RefQuery	$\approx 1$	Yes	$4.145+K\cdot 0.165$

TABLE IX: Scalability with number of appliances

K

\Delta W

is the additional storage per added appliance. “Incremental” means adding an appliance without updating shared parameters. Inference scaling is reported in MFLOPs/(mains window) as

b+K\cdot m

, where

b

is the

K

-independent compute (shared cost) and

m

is the additional MFLOPs per added appliance.

Non-Intrusive Load Monitoring

Non-intrusive load monitoring estimates individual appliance use from a household’s combined electricity measurements.

See all 12 papers

Ask for Any Appliance: A Prompt-Programmable Foundation Model for Non-Intrusive Load Monitoring
Sep 19, 2026Xudong Wang, Jiacheng Cui, Junyu Xue +2Non-Intrusive Load Monitoring Time-Series Foundation Models
The Chinese University of Hong Kong, Shenzhen · The Hong Kong University of Science and Technology (Guangzhou) · Southern University of Science and Technology
Energy Injection Identification enabled Disaggregation with Deep Multi-Task Learning
Aug 20, 2025Xudong Wang, Guoming Tang, Junyu Xue +3Non-Intrusive Load Monitoring Multi-Task Learning
The Chinese University of Hong Kong, Shenzhen · The Hong Kong University of Science and Technology (Guangzhou) · Southern University of Science and Technology +1
Multi-Appliance Non-Intrusive Load Monitoring via Label-Preserving Aggregate Recomposition and Prediction Consistency
Sep 16, 2026Jiangfeng Liu, Yanfang FanNon-Intrusive Load Monitoring Data Augmentation
School of Electrical Engineering, Xinjiang University, Urumqi, Xinjiang 830047, China

Sep 19, 2026J/K move · Enter open · S save

Ask for Any Appliance: A Prompt-Programmable Foundation Model for Non-Intrusive Load Monitoring

Xudong Wang, Jiacheng Cui, Junyu Xue, Tongxin Li+1

The Chinese University of Hong Kong, Shenzhen · The Hong Kong University of Science and Technology (Guangzhou) · Southern University of Science and Technology

Non-Intrusive Load Monitoring Time-Series Foundation Models

Open paper AskPDF arXiv

Non-intrusive load monitoring (NILM) estimates appliance-level consumption from a whole-home meter, but appliance-specific models and fixed output inventories make coverage costly to extend. We present FM4NILM (Foundation Model for NILM), a single prompt-programmable model that estimates a requested appliance's power trajectory from aggregate measurements, a natural-language description, and optional activation exemplars. A lightweight cadence-aware transformer is pretrained by masked reconstruction on 645k sequences from seven public corpora spanning 1-60 s sampling intervals, then aligned with appliance requests using observation-masked losses for partially labeled households. A Bernoulli-lognormal decoder separates activity detection from conditional power estimation. On held-out households and time periods from REDD, UK-DALE, and REFIT, one frozen text-prompted model serves twelve appliance-corpus requests, achieving 0.556 event F1, 0.625 AUPRC, and the lowest active-window MAE (251.8 W) among seven appliance-specific baselines. Streaming score aggregation raises event F1 to 0.582 with a 60 s aggregation delay. In a separate category-held-out evaluation, adding ten activation exemplars raises microwave AUPRC from 0.132 to 0.214 without parameter updates. Input-intervention ablations probe the model's dependence on appliance requests and aggregate measurements. These results demonstrate competitive disaggregation with one shared model and support extending appliance coverage through prompts and examples rather than additional specialist networks.

Abstract

Non-Intrusive Load Monitoring

Explore similar work

Ask for Any Appliance: A Prompt-Programmable Foundation Model for Non-Intrusive Load Monitoring

Energy Injection Identification enabled Disaggregation with Deep Multi-Task Learning

Multi-Appliance Non-Intrusive Load Monitoring via Label-Preserving Aggregate Recomposition and Prediction Consistency