acceptodds
Withdrawn from ICLR 2027

Appliance Signature Space with One For All NILM

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

NNon-intrusive load monitoring (NILM) estimates the power drawn by individual appliances from a single household meter. Deep NILM models are usually trained one per appliance, so a smart meter that tracks many appliances must store many networks. We study one-for-all NILM (OFAN), in which a single transformer disaggregates whichever appliance it is conditioned on through a token appended to its input sequence, taken either from a learned embedding or from an encoder applied to a representative activation waveform. We add this conditioning to two models: BERT4NILM and NILMformer, and evaluate it on UK-DALE and REFIT over five seeds with paired confidence intervals. For five appliances, one conditioned BERT4NILM model has 1.9M parameters compared to 9.7M for five separate models, and each additional appliance adds a single 256-dimensional embedding. The cost of sharing one model depends on the backbone and the dataset: on UK-DALE, pooled mean absolute error (MAE) rises by 12.8% for NILMformer and by 24.6% for BERT4NILM, while on the REFIT datasets, a single BERT4NILM model lowers MAE by 5.6%. Conditioning on activation waveforms is less accurate than learned embeddings on UK-DALE; however, it transfers what it has learned: applied to the other dataset without retraining and conditioned on the appliance signature it learned in training, a model reaches a mean F1 of over 101 sub-metered channels, above a random detector on every channel and above an always-on detector on 88 of them, and F1 higher than the same model conditioned on a query built in the target home (95% CI ); appliance types absent from training are not detected. Finally, we read the learned embedding table as a proxy for Hart's appliance signature space and find structure in it that only training can explain: in all four model–dataset pairs, the kettle and microwave, the two appliances with similar short high-power cycles, form the closest pair, and in three of them, they trade accuracy under joint training.

Primary area:
learning on time series and dynamical systems
Keywords:
non-intrusive load monitoring, energy disaggregation, conditional transformers, appliance embeddings, representation learning, zero-shot transfer, appliance signature space

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_h2KkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes OFAN, a conditional NILM framework that appends an appliance embedding to the input of a transformer backbone so that one shared model can disaggregate multiple appliances. OFAN is instantiated on BERT4NILM and NILMformer and evaluated on UK-DALE and REFIT with five seeds.OFAN is literally a single conditioned model plus a small embedding table, reducing stored parameters for five appliances from 9.7M to about 1.9M for BERT4NILM.

Critical strengths
  1. This paper presents an extensive empirical study. It separates the effects of the conditioning token, joint training, and query type, and uses paired confidence intervals over five seeds.
  2. OFAN shows useful transfer learning analysis. The zero-shot transfer experiment, conditioning on the training signature without target-home data, is a practically interesting result for the NILM community.
Critical weaknesses
  1. This paper shows limited methodological novelty. NILM is a mature field, and appending a learned embedding to a transformer input is a standard conditioning mechanism closely related to FiLM and QueryNILM. The paper does not introduce a new architecture or learning algorithm.
  2. The evaluation is too narrow to validate scalability. The main experiments cover only four or five appliance categories from two UK datasets, with one held-out test house for UK-DALE and two for REFIT.
  3. The deployment motivation is storage on constrained smart meters, but the paper counts parameters only and does not measure on-device memory or latency, which may be unacceptable in deployment.
Questions for authors
  1. Given QueryNILM’s substantially smaller parameter count and comparable or better index-query performance on REFIT, what practical advantage does OFAN offer?
Reviewer_dqsHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes One For All NILM (OFAN), a unified Transformer-based framework for multi-appliance non-intrusive load monitoring. By introducing appliance-conditioned embedding tokens into BERT4NILM and NILMformer, OFAN enables a single shared model to disaggregate different appliances. The authors investigate learnable index embeddings and activation-based embeddings. Experiments on UK-DALE and REFIT demonstrate substantial parameter reduction, although prediction accuracy varies across datasets and architectures. The paper also analyzes learned appliance representations and evaluates cross-dataset transfer.

Critical strengths

1.The paper addresses the practical problem of reducing model storage requirements for multi-appliance NILM systems, which is relevant to resource-constrained smart meters.

2.The analysis of appliance embeddings and cross-dataset transfer provides interesting insights into shared representations and appliance similarities.

Critical weaknesses

1.The proposed method primarily appends an appliance embedding token to existing Transformer architectures without introducing new attention mechanisms or optimization objectives. Similar conditional multi-appliance disaggregation has already been explored by QueryNILM. The paper should better clarify its methodological novelty beyond adapting existing conditioning techniques to Transformers.

2.Although the paper compares OFAN with QueryNILM, other relevant multi-appliance approaches, such as MA-CNN, FCNdAE, and RefQuery, are not evaluated. Moreover, QueryNILM achieves better accuracy on REFIT with index queries while requiring substantially fewer parameters. This raises questions about the advantages of OFAN over existing methods.

3.The experiments measure parameter reduction without evaluating inference latency, computational cost, or memory consumption on actual smart meter hardware. Since each appliance still requires a separate forward pass, the practical efficiency benefits remain unclear.

4.The proposed unified model introduces a noticeable accuracy trade-off. On UK-DALE, joint training increases MAE by 12.8%–24.6% compared to individual appliance models, with particularly significant degradation for certain appliances. This raises concerns about whether the shared architecture can effectively balance parameter efficiency and disaggregation accuracy.

5.The cross-dataset experiments primarily evaluate appliance types already seen during training, while the model performs poorly on unseen appliance categories. Therefore, the results do not sufficiently demonstrate the model's ability to generalize to novel appliance types, limiting its applicability to real-world scenarios.

Questions for authors

Please refer to the Critical Weaknesses section above.

Reviewer_5NX4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes OFAN, a conditional transformer formulation for non-intrusive load monitoring. A single BERT4NILM or NILMformer backbone receives either a learned appliance identifier embedding or a waveform-derived activation token and predicts the requested appliance trace. Experiments on UK-DALE and REFIT compare independent and jointly trained models, analyze learned token similarity, and test cross-dataset performance.

Critical strengths
  • The paper poses a concrete deployment question: reducing the storage cost of maintaining one network per appliance. It evaluates two backbones, two conditioning schemes, and reports results across five seeds rather than presenting a single selected run.
  • The results are more candid than usual: joint training hurts on UK-DALE, and the activation-token experiment explicitly shows that appliance types absent from training are not detected.
Critical weaknesses
  • The related-work treatment is materially incomplete for the paper’s central representational claim. The omission makes both the novelty claim and the embedding interpretation inadequately supported.
  • The headline scalability/generalization claim is overstated. The identifier embedding cannot add an unseen appliance without training a new embedding, and the waveform conditioned model fails on appliance types absent from training. The reported “zero-shot” transfer instead reuses a signature for an appliance type already seen in source training and evaluates it in another dataset; it is cross-dataset transfer of known classes, not open-world appliance disaggregation. On UK-DALE, the shared model also incurs large pooled MAE increases (12.8% and 24.6% for the two backbones). The paper therefore demonstrates a storage-performance trade-off for a fixed known set of appliances, not one-for-all NILM in the claimed general sense.
Questions for authors
  1. Please add and compare against recent NILM representation-learning/disentanglement work, and clarify precisely what is novel beyond existing query-conditioned NILM.

  2. Can the authors evaluate an actual appliance-registration setting: a new appliance class, no output-head change, a bounded amount of its signature data, and comparison with retraining or adaptation baselines? The present results show that the proposed waveform query does not solve this problem.

  3. Please report cross-house splits, the exact relationship between source and target appliance channels, inference latency/peak memory on representative edge hardware, and energy or aggregate-consistency metrics. Parameter count alone does not establish deployability on a smart meter.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.