Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Cluster Aggregated GAN (CAG): A Cluster-Based Hybrid Model for Appliance Pattern Generation

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that decomposing appliance power traces into behavioral clusters and training a dedicated generator per cluster produces synthetic load data that is more realistic, more diverse, and more stable to train than any single mo

desk verdict Good idea, poor verification: the diversity metrics are scored backwards and the coverage test is circular; worth a referee but not acceptance. read the letter →

arxiv 2512.22287 v3 pith:66OX64Z7 submitted 2025-12-25 cs.LG cs.AI

Zikun Guo , Adeyinka.P. Adedigba , Rammohan Mallipeddi This is my paper · ORCID
classification cs.LGcs.AI
keywords generativeadversarialnetworksnon-intrusiveloadmonitoringappliancepatterngenerationclusteringLSTMsynthetictimeseriesUVICdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the reason GAN-based synthetic appliance data falls short is that a single generator tries to model fundamentally different behaviors at once. To fix that, CAG first routes each device to an intermittent or continuous branch via three cheap statistics, then clusters the intermittent segments by shape and trains a dedicated small GAN on each cluster, while the continuous branch uses an LSTM generator on a compressed sequence with a shared discriminator. On the UVIC dataset, the authors report that CAG beats CNN, LSTM, RNN, and WaveGAN baselines on realism, diversity, and training stability, and that removing the clustering stage sharply degrades performance. If correct, this gives NILM researchers a privacy-preserving way to generate appliance data that preserves rare operational modes and is more interpretable.

What carries the argument

The load-bearing mechanism is the cluster-aggregated routing loop: a three-feature classifier (R0, pnz, σ²Δ) sends each device to an intermittent or continuous branch; intermittent traces are projected into a shape-feature space (moments, trend, dominant frequency, morphological counts, roughness, energy, and 20 shape samples), then partitioned by K-means whose K is chosen by maximum silhouette score; each cluster trains an independent lightweight CNN-GAN, so the generator sees a unimodal distribution; a shared discriminator pools adversarial signal across branches; continuous traces are compressed by uniform averaging (factor F) and reconstructed by block replication, keeping the LSTM-GAN o

What would settle it

Recompute cluster coverage and cluster JS divergence using clusters that are fixed in advance and independent of CAG—for instance, clusters learned from a completely separate dataset or defined by hand-labeled operational modes—then retrain the CNN/LSTM/RNN/WaveGAN baselines under the same protocol; if CAG's diversity advantage over the best baseline shrinks or disappears, the headline result is at least in part an artifact of scoring on CAG's own cluster partition. Also, train the strongest baseline (e.g., WaveGAN) with the same per-cluster generator allocation; if it matches CAG's numbers, t

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that aligning model structure with behavioral structure—rather than the appliance label—is what drives generative quality. The CAG pipeline routes each device by three lightweight statistics: initial zero inactivity, nonzero occupancy, and smoothed derivative variance. Intermittent devices are segmented, shape-featurized, and K-means clustered with silhouette-selected K; each cluster trains an independent CNN-based GAN so that rare operational modes receive dedicated capacity. Continuous devices are uniformly downsampled, modeled by an LSTM-GAN, and reconstructed by block replication, with a shared discriminator enforcing a common realism standa

Load-bearing premise

The diversity metrics that make CAG look best are computed on the same K-means clusters that CAG's generators are trained to reproduce, so the comparison implicitly assumes that partition is a neutral benchmark for every baseline model.

Editorial extensions

If this is right

  • Synthetic appliance loads generated this way should preserve rare operational modes (e.g., late-night microwave cycles) that a monolithic GAN tends to collapse, improving downstream NILM training data.
  • Because each cluster corresponds to an interpretable usage pattern, practitioners can inspect or steer which modes are generated, aiding debugging and domain-constrained generation.
  • The routing rule is label-free and portable: new appliance traces can be assigned to branches or even to existing clusters without retraining the classifier.
  • If the framework generalizes, the cluster-then-generate design applies to any multimodal time-series domain, not just energy data.
  • Training stability gains mean fewer epochs of oscillation and more predictable convergence, which lowers the cost of deploying GAN-based data synthesis.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the diversity metrics (cluster coverage and cluster JS divergence) are computed against the K-means partition that CAG itself generates, the reported diversity advantage over baselines could partly reflect that CAG is being scored on its own home grid; an independent fixed taxonomy (e.g., clusters learned from a different household set or hand-labeled modes) would test that.
  • The paper's own Table 4 shows the routing rule sometimes mislabels devices (e.g., CoffeeMaker 'continuous', IMac 'intermittent'), yet the clustered branch still helps those devices; this hints that clustering, not the routing accuracy, is the main driver—a testable separation.
  • A natural extension: condition the per-cluster generators on cluster identity or usage context (time of day, occupancy) rather than generating clusters independently, which could improve inter-cluster coherence for downstream NILM.
  • The silhouette-based K selection treats each device separately; an alternative is a global, budget-constrained cluster count across all devices, which would speak to scalability and the paper's stated interpretability goal.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes Cluster Aggregated GAN (CAG), a two-branch generative framework for appliance load pattern synthesis. Intermittent appliances are segmented, clustered by shape features, and each cluster receives a dedicated CNN-based GAN; continuous appliances are downsampled and modeled by an LSTM-GAN. On the UVIC dataset, CAG is compared to CNN, LSTM, RNN, and WaveGAN baselines using realism metrics (ME, Std, Fid, Per, Feature FID), diversity metrics (Div, CC, CJ), and qualitative convergence behavior. The paper claims that CAG consistently outperforms baselines across realism, diversity, and training stability, and that clustering as an active generative component improves interpretability and scalability.

Significance. If the empirical claims held, the core idea of routing appliance traces into per-cluster generators for intermittent devices plus an LSTM branch for continuous devices would be a useful contribution to synthetic NILM data generation, especially for allocating capacity to rare operational modes. The paper also contains a useful ablation study (Table 5) and a cluster-strategy sweep (Table 4). However, the evaluation contains a diversity metric that is scored in the wrong direction and a cluster-based comparison that is partly circular; as a result, the central claim of consistent superiority is not supported by the evidence presented.

major comments (4)
  1. [§4.3.1, Table 2, Table 5] Cluster JS (CJ) is defined as a Jensen–Shannon divergence and the text states 'Smaller values indicate balanced sampling across behavioural clusters.' Table 2 nevertheless displays CJ with an upward arrow and reports CAG's 6.57e-1 as part of 'the largest diversity metrics,' while all baselines are near 0.346-0.372. Under the paper's own definition, CAG has the least balanced cluster distribution, not the best. Table 5 repeats the error: for Water Cooler, the text calls an increase from 3.10e-1 to 6.45e-1 an 'improvement' of cluster JS. This reverses the meaning of the metric and invalidates the diversity dimension of the headline claim.
  2. [§3.3, §4.3.1] CC and CJ are computed from K-means clusters of the real segments, and §3.3 builds CAG's per-cluster generators on exactly this partition. A method given one generator per cluster is structurally favored to place samples in all clusters, whereas the monolithic baselines receive no cluster information. No protocol is described for assigning generated segments to clusters, and it is unclear why CAG's CC is only 0.303 instead of near 1 if each cluster has a dedicated generator. This comparison is not neutral; independent clustering or equivalently conditioned baselines is needed before diversity gains can be attributed to better generation.
  3. [§4.3, Table 2] Feature FID depends on an unspecified feature extractor Φ. The reported values (CAG 5.82e16, baselines roughly 1e16-1e17) are not interpretable without knowing Φ's dimension and scale; a Fréchet distance with a squared covariance trace can be arbitrarily large. No definition or reference for Φ is given, so the realism comparison on this metric cannot be assessed. The authors should specify the extractor and report normalized or per-dimension values.
  4. [§3.3, §3.4, Figure 2] The training setup is described inconsistently. §3.3 says each cluster trains 'an independent lightweight fully connected GAN' with its own discriminator, while §3.4 says 'CAG uses a shared discriminator D across all generated outputs' and Figure 2 also shows one shared discriminator. The number of discriminators and the exact combined loss are central to the method; as written, the architecture cannot be reproduced unambiguously.
minor comments (4)
  1. [§5.2] The text says the results 'empirically confirm the theoretical convergence analysis described earlier,' but no theoretical convergence analysis appears in Sections 1-4.
  2. [General] Tables 2 and 3 report single numbers without error bars, multiple seeds, or significance tests; claims of 'consistent' superiority would be stronger with repeated-run statistics.
  3. [§5.4, Table 4] The detected types in Table 4 contradict examples in the introduction: CoffeeMaker is detected as continuous, IMac as intermittent, and Refrigerator as intermittent. This is possible under the heuristic, but the discrepancy should be acknowledged and explained.
  4. [Abstract] The abstract contains a typographical error: 'trainingstabilitythroughsequencecompression' should be separated into words.

Circularity Check

1 steps flagged · score 5.0 of 10

Partial circularity in the diversity evaluation: cluster coverage uses the same K-means partition that defines CAG's per-cluster generators, so the coverage advantage is partly built in; the cluster-JS column is also scored in the wrong direction.

  1. fitted input called prediction [§3.3 (Per-Cluster GAN Training) and §4.3.1 (Cluster Coverage); used in Table 2]
    "§3.3: 'For cluster k, the segment dataset is D(k)={s(j):c(j)=k}... Each cluster dataset D(k) trains an independent lightweight fully connected GAN.' §4.3.1: 'Given K behavioural clusters extracted from real data, let n(g)k be the number of generated samples assigned to cluster k. Coverage is defined as Clus.Cov.=1/K Σ_k 1[n(g)k>0].'"

    CC tests whether generated samples land in each of the K real-data clusters, but CAG is constructed from those same clusters: one generator is trained per cluster on exactly the real segments assigned to it. A per-cluster generator is therefore structurally meant to populate the cluster it was built for, so high CC mainly re-states the training partition rather than measuring independent diversity. Monolithic baselines are not given this partition, making Table 2's CC comparison non-neutral. The cluster-JS column adds a further inconsistency: §4.3.1 states 'Smaller values indicate balanced sampling,' yet Table 2 marks CJ with ↑ and counts CAG's larger 6.57e-1 as an advantage. The realism metrics remain independent, so the paper is not wholly circular.

full rationale

The architecture itself is not derived from the evaluation metrics: routing, clustering, and adversarial losses are standard and self-contained. The realism results (ME, Std, Fid, Per, Fea) provide independent evidence for CAG's fidelity claims. However, the headline diversity claim relies heavily on cluster coverage, and CC is evaluated on the same K-means partition that defines the per-cluster generator structure. That part of the comparison is circular in a partial sense: the model is built to produce each cluster, and the metric simply checks whether it did. The cluster-JS 'balance' metric is scored in the opposite direction in Table 2, which further undermines the diversity claim, though this is a correctness issue rather than circularity. Self-citations [15,16] are background and not load-bearing. Overall, the central realism claim survives, but the diversity component is partly forced by construction, giving a score of 5.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claim rests on many hand-set thresholds and data-dependent design choices (segment length 436, routing thresholds, cluster counts), none of which are given error bars or validated on held-out data. The evaluation also relies on cluster-derived metrics aligned with the method. No new physical entities are introduced.

free parameters (9)
  • Classification threshold T0 = 100
    Prefix length for initial-inactivity detection; set in §3.2 based on empirical analysis of UVIC.
  • Occupancy threshold rho = 0.7
    Nonzero-occupancy split between continuous and intermittent; set on UVIC (§3.2).
  • Derivative-variance threshold tau = 0.1
    Volatility threshold for routing; set on UVIC (§3.2).
  • Segment length L = 436
    Fixed-length non-overlapping window for intermittent branch; acknowledged in conclusion as a limitation.
  • Downsampling factor F = not reported (≤1000 points)
    Uniform averaging factor Eq. 8; the actual F used is not given.
  • Max cluster count kappa = not reported
    Upper bound for silhouette-based K selection (§3.3); value not stated.
  • Per-device cluster count K = 2–10 (Table 4)
    Selected by silhouette on the same UVIC data used for evaluation; determines number of generators.
  • Square-wave separation gamma = not reported
    Threshold in Eq. 10 for detecting bistable devices.
  • Spike quantile q and window S = not reported
    Parameters for spiky-device spike extraction (§3.4).
assumptions (5)
  • domain assumption K-means on handcrafted shape features yields behavioral clusters that correspond to real operational modes.
    Central to the cluster-aggregation claim; no external validation of cluster semantics.
  • ad hoc to paper The heuristic thresholds (T0=100, rho=0.7, tau=0.1) correctly separate continuous and intermittent appliances.
    Set on UVIC; Table 4 mislabels devices that Figure 3 calls intermittent (CoffeeMaker, Microwave, Printer).
  • standard math GAN training with logistic loss and per-cluster unimodal targets converges and avoids mode collapse.
    Standard GAN assumption; paper offers no formal proof despite referencing 'theoretical convergence analysis'.
  • domain assumption Uniform downsampling Eq. 8 and block replication Eq. 9 preserve the temporal structure of continuous devices.
    Piecewise-constant reconstruction may distort transient behavior; not validated separately.
  • ad hoc to paper Cluster coverage and cluster JS computed on the same clusters used to train CAG are an unbiased comparison across all baselines.
    Structurally favors a per-cluster generator; flagged as circular.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cluster Aggregated GAN (CAG): A Cluster-Based Hybrid Model for Appliance Pattern Generation." pith.science (2026). https://pith.science/paper/66OX64Z7

@misc{pith2026251222287,
  author       = {Pith},
  title        = {Pith review of: Cluster Aggregated GAN (CAG): A Cluster-Based Hybrid Model for Appliance Pattern Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66OX64Z7}},
  note         = {Machine review of arXiv:2512.22287}
}
read the original abstract

Synthetic appliance data are essential for developing non-intrusive load monitoring algorithms and enabling privacy preserving energy research, yet the scarcity of labeled datasets remains a significant barrier. Recent GAN-based methods have demonstrated the feasibility of synthesizing load patterns, but most existing approaches treat all devices uniformly within a single model, neglecting the behavioral differences between intermittent and continuous appliances and resulting in unstable training and limited output fidelity. To address these limitations, we propose the Cluster Aggregated GAN framework, a hybrid generative approach that routes each appliance to a specialized branch based on its behavioral characteristics. For intermittent appliances, a clustering module groups similar activation patterns and allocates dedicated generators for each cluster, ensuring that both common and rare operational modes receive adequate modeling capacity. Continuous appliances follow a separate branch that employs an LSTM-based generator to capture gradual temporal evolution while maintaining training stability through sequence compression. Extensive experiments on the UVIC smart plug dataset demonstrate that the proposed framework consistently outperforms baseline methods across metrics measuring realism, diversity, and training stability, and that integrating clustering as an active generative component substantially improves both interpretability and scalability. These findings establish the proposed framework as an effective approach for synthetic load generation in non-intrusive load monitoring research.

Figures

Figures reproduced from arXiv: 2512.22287 by the authors.

Figure 1
Figure 1. Overall workflow of the proposed Cluster Aggregated GAN framework for appliance load generation. Appliance traces are routed into intermittent or continuous branches by a lightweight classifier. Intermittent streams are clustered into patterns and modeled by dedicated CNN based generators for each cluster. Continuous streams are modeled by an LSTM based generator. A shared discriminator enforces realism across branc… view at source ↗
Figure 2
Figure 2. Cluster Aggregation GAN (CAG) overview. UVIC traces are routed by a lightweight classifier into intermittent and continuous devices. Intermittent streams are segmented, clustered, and generated via CNN based GANs assigned to each cluster. Continuous devices are downsampled and modeled by an LSTM based GAN. A shared discriminator 𝐷 judges realism for all branches, promoting consistency across heterogeneous appliance … view at source ↗
Figure 3
Figure 3. Visualization of appliance load profiles in the dataset. Representative time series power consumption patterns for multiple household and office appliances from the UVIC dataset. Each subplot shows the active power over time for a specific device, illustrating the diversity of load behaviors across appliance types. Intermittent devices such as the CoffeeMaker, Microwave, and Printer exhibit short, sporadic activatio… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Example output visualization of the proposed CAG framework. The generated appliance load patterns closely resemble the real data, demonstrating the effectiveness of the cluster learning strategy and adaptive routing mechanism [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Convergence comparison of different generative models. The proposed CAG framework exhibits the most stable and consistent convergence behavior, with the generator and discriminator losses gradually approaching equilibrium. In contrast, baseline models such as CNN-base,…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CADET: Physics-Grounded Causal Auditing and Training-Free Deconfounding of End-to-End Driving Planners

    cs.RO 2026-06 unverdicted novelty 5.0 of 10

    CADET is a training-free framework for auditing, benchmarking, and repairing spurious correlations in pretrained end-to-end autonomous driving planners using physics-grounded causal methods.

Reference graph

Works this paper leans on

40 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    Privacyprotectionofsyntheticsmartgriddatasimulatedviagenerativeadversarialnetworks,in:Proceedings of the 20th International Conference on Security and Cryptography (SECRYPT), pp

    Adewole,K.S.,Torra,V.,2023. Privacyprotectionofsyntheticsmartgriddatasimulatedviagenerativeadversarialnetworks,in:Proceedings of the 20th International Conference on Security and Cryptography (SECRYPT), pp. 279–286

    2023

  2. [2]

    NILM applications: Literature review of learning approaches, recent developments and challenges

    Angelis, G.F., et al., 2022. NILM applications: Literature review of learning approaches, recent developments and challenges. Energy and Buildings 261, 111951. doi:10.1016/j.enbuild.2022.111951

    arXiv 2022

  3. [3]

    Wasserstein GAN, in: Proceedings of the 34th International Conference on Machine Learning (ICML), pp

    Arjovsky, M., Chintala, S., Bottou, L., 2017. Wasserstein GAN, in: Proceedings of the 34th International Conference on Machine Learning (ICML), pp. 214–223

    2017

  4. [4]

    Algan: Time series anomaly detection with adjusted-lstm gan

    Bashar, M.A., Nayak, R., 2023. Algan: Time series anomaly detection with adjusted-lstm gan. arXiv preprint arXiv:2308.06663

    arXiv 2023

  5. [5]

    NILMTK: An open source toolkit for energy disaggregation, in: Proceedings of the 5th International Conference on Future Energy Systems (ACM e-Energy), pp

    Batra, N., Kelly, J., Parson, O., Bergés, M., 2014. NILMTK: An open source toolkit for energy disaggregation, in: Proceedings of the 5th International Conference on Future Energy Systems (ACM e-Energy), pp. 265–276

    2014

  6. [6]

    Review of internet of things in electric power and energy systems

    Bedi, G., Aujla, G.S., Tanwar, S., Kumar, R., Kumar, N., 2018. Review of internet of things in electric power and energy systems. IEEE Internet of Things Journal 5, 847–870. doi:10.1109/JIOT.2018.2802704

    arXiv 2018

  7. [7]

    Ambal: Appliance modeling by aggregation of load segments, in: Proceedings of the 8th International Conference on Future Energy Systems (e-Energy), pp

    Buneeva, N., Reinhardt, A., 2017. Ambal: Appliance modeling by aggregation of load segments, in: Proceedings of the 8th International Conference on Future Energy Systems (e-Energy), pp. 287–292

    2017

  8. [8]

    Residential load modeling with generative adversarial networks

    Castangia, M., et al., 2025. Residential load modeling with generative adversarial networks. IEEE Transactions on Sustainable Computing Early Access

    2025

Show all 40 references
  1. [9]

    SmartSim: A device-accurate smart home simulator for energy analytics, in: Proceedings of the IEEE International Conference on Smart Grid Communications (SmartGridComm 2016), pp

    Chen, D., Irwin, D., Shenoy, P., 2016. SmartSim: A device-accurate smart home simulator for energy analytics, in: Proceedings of the IEEE International Conference on Smart Grid Communications (SmartGridComm 2016), pp. 686–692

    2016

  2. [10]

    A data-driven approach for generating synthetic load patterns and usage habits

    El Kababji, S., Srikantha, P., 2020. A data-driven approach for generating synthetic load patterns and usage habits. IEEE Transactions on Smart Grid 11, 4984–4995

    2020

  3. [11]

    Real-valued (medical) time series generation with recurrent conditional gans

    Esteban, C., Hyland, S.L., Rätsch, G., 2017. Real-valued (medical) time series generation with recurrent conditional gans. arXiv preprint arXiv:1706.02633

    2017 arXiv

  4. [12]

    Asurveyonnon-intrusiveloadmonitoring

    Faustine,W.,Suchithra,N.,Mushi,L.,etal.,2017. Asurveyonnon-intrusiveloadmonitoring. Energies10,957. doi:10.3390/en10070957

    2017 doi

  5. [13]

    Gkoutroumpi, C., Gkalinikis, N.V., Vrakas, D., 2024. SGAN: Appliance signatures data generation for NILM applications using GANs, in: Proceedings of the 12th Computing Conference (Science and Information Conference, SAI 2024), Springer. pp. 325–339. doi:10.1007/ 978-3-031-62269-4\_23

    2024

  6. [14]

    Generative adversarial networks

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y., 2020. Generative adversarial networks. Communications of the ACM 63, 139–144

    2020

  7. [15]

    Cluster-aggregated transformer: Enhancing lightweight parameter models

    Guo, Z., Adedigba, A.P., Mallipeddi, R., 2025a. Cluster-aggregated transformer: Enhancing lightweight parameter models. Engineering Applications of Artificial Intelligence 159, 111468. doi:10.1016/j.engappai.2025.111468

    2025

  8. [16]

    Dynamic tanh reinforcement learning: A normalization-free transformer for open traveling salesman problem optimization

    Guo, Z., Adedigba, A.P., Mallipeddi, R., Lee, H., 2025b. Dynamic tanh reinforcement learning: A normalization-free transformer for open traveling salesman problem optimization. Journal of the Institute of Control, Robotics and Systems Conference Proceedings , 845–846

  9. [17]

    Guo, Z., Kavuri, S., Lee, J., Lee, M., 2023. IDS-Extract: Downsizing deep learning model for question and answering, in: Proceedings of the 2023 International Conference on Electronics, Information, and Communication (ICEIC), pp. 1–5. doi:10.1109/ICEIC57457.2023. 10049915

    2023

  10. [18]

    TraceGAN: Synthesizing appliance power signatures using generative adversarial networks

    Harell, A., Jones, R., Makonin, S., Bajić, I.V., 2021. TraceGAN: Synthesizing appliance power signatures using generative adversarial networks. IEEE Transactions on Smart Grid 12, 4553–4563. doi:10.1109/TSG.2021.3078695. First Author et al.:Preprint submitted to ElsevierPage 1...

    2021

  11. [19]

    Nonintrusive appliance load monitoring

    Hart, G.W., 1992. Nonintrusive appliance load monitoring. Proceedings of the IEEE 80, 1870–1891. doi:10.1109/5.192069

    1992 doi

  12. [20]

    Tcgan:Convolutionalgenerativeadversarialnetworkfortimeseriesclassificationandclustering

    Huang,F.,Deng,Y.,2023. Tcgan:Convolutionalgenerativeadversarialnetworkfortimeseriesclassificationandclustering. NeuralNetworks 165, 868–883

    2023

  13. [21]

    SIDED: Synthetic industrial dataset for energy disaggregation via digital twins

    Internò, C., et al., 2025. SIDED: Synthetic industrial dataset for energy disaggregation via digital twins. arXiv preprint arXiv:2506.20525. URL:https://arxiv.org/abs/2506.20525

    2025

  14. [22]

    Sensors 24, 5532

    Jaradat,A.,Al-Sareh,M.,Al-Ghazo,S.,etal.,2024.HYDROSAFE:Ahybriddeterministic-probabilisticmodelforsyntheticapplianceprofiles generation. Sensors 24, 5532. doi:10.3390/s24155532

    2024 doi

  15. [23]

    HiFAKES: Synthetic high-frequency NILM data for NILM models diagnostics and generalization testing

    Kamyshev, I., Moghimian, S., Ouerdane, H., 2025. HiFAKES: Synthetic high-frequency NILM data for NILM models diagnostics and generalization testing. arXiv preprint arXiv:2409.00062 URL:https://arxiv.org/abs/2409.00062

    2025 arXiv

  16. [24]

    Kelly, J., Knottenbelt, W., 2015. Neural NILM: Deep neural networks applied to energy disaggregation, in: Proceedings of the 2nd ACM International Conference on Embedded Systems for Energy-Efficient Built Environments (BuildSys), pp. 55–64

    2015

  17. [25]

    Auto-encoding variational bayes, in: Proceedings of the International Conference on Learning Representa- tions (ICLR)

    Kingma, D.P., Welling, M., 2014. Auto-encoding variational bayes, in: Proceedings of the International Conference on Learning Representa- tions (ICLR)

    2014

  18. [26]

    SynD: A synthetic energy dataset for non-intrusive load monitoring in households

    Klemenjak, C., Kovatsch, C., Herold, M., Elmenreich, W., 2020. SynD: A synthetic energy dataset for non-intrusive load monitoring in households. Scientific Data 7, 108. doi:10.1038/s41597-020-0434-6

    2020 doi

  19. [27]

    LearningandgeneratingdiverseresidentialloadpatternsusingGANwithweakly-supervisedtrainingandweight selection

    Liang,X.,Wang,H.,2025. LearningandgeneratingdiverseresidentialloadpatternsusingGANwithweakly-supervisedtrainingandweight selection. IEEE Transactions on Consumer Electronics 71, 2837–2848. doi:10.1109/TCE.2025.3563272

    2025

  20. [28]

    UsingGANsforsharingnetworkedtimeseriesdata,in:ProceedingsoftheACMInternetMeasurement Conference (IMC), pp

    Lin,Z.,Han,J.,Zhu,Q.,etal.,2020. UsingGANsforsharingnetworkedtimeseriesdata,in:ProceedingsoftheACMInternetMeasurement Conference (IMC), pp. 464–481

    2020

  21. [29]

    VA-creator: A virtual appliance creator based on adaptive neural networks to generate synthetic power consumption patterns

    Meiser, M., Düppe, B., Zinnikus, I., 2024. VA-creator: A virtual appliance creator based on adaptive neural networks to generate synthetic power consumption patterns. Energy and AI 8, 100205

    2024

  22. [30]

    Conditional generative adversarial nets

    Mirza, M., Osindero, S., 2014. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784. URL:https://arxiv.org/abs/ 1411.1784

    2014 arXiv

  23. [31]

    ClusterGAN: Latent space clustering in generative adversarial networks, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Mukherjee, S., Asnani, H., Alfonzo, E., Kannan, S., 2019. ClusterGAN: Latent space clustering in generative adversarial networks, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 4610–4617

    2019

  24. [32]

    Machinelearningapproachesfornon-intrusiveloadmonitoring:fromqualitativetoquantitativecomparison

    Nalmpantis,C.,Vrakas,D.,2019. Machinelearningapproachesfornon-intrusiveloadmonitoring:fromqualitativetoquantitativecomparison. Artificial Intelligence Review 52, 217–243. doi:10.1007/s10462-018-9650-2

    2019 doi

  25. [33]

    Conditional image synthesis with auxiliary classifier GANs, in: Proceedings of the 34th International Conference on Machine Learning (ICML), pp

    Odena, A., Olah, C., Shlens, J., 2017. Conditional image synthesis with auxiliary classifier GANs, in: Proceedings of the 34th International Conference on Machine Learning (ICML), pp. 2642–2651

    2017

  26. [34]

    Conditional GAN for time series generation

    Smith, K.E., Smith, A.O., 2020. Conditional GAN for time series generation. arXiv preprint arXiv:2006.16477. URL:https://arxiv. org/abs/2006.16477

    2020 arXiv

  27. [35]

    ELECTRIcity: An efficient transformer for non-intrusive load monitoring

    Sykiotis, S., Kaselimi, M., Doulamis, A., Doulamis, N., 2022. ELECTRIcity: An efficient transformer for non-intrusive load monitoring. Sensors 22, 2926. doi:10.3390/s22082926

    2022 doi

  28. [36]

    STNILM: Switch transformer-based NILM adaptive to households

    Varanasi, L.S., Karri, S.P.K., 2023. STNILM: Switch transformer-based NILM adaptive to households. Energy and Buildings 275, 113249

    2023

  29. [37]

    DPP-GAN: A decentralized and privacy-preserving GAN system for collaborative smart meter data generation

    Wang, X., et al., 2021. DPP-GAN: A decentralized and privacy-preserving GAN system for collaborative smart meter data generation. arXiv preprint arXiv:2110.12345. URL:https://arxiv.org/abs/2110.12345

    2021 arXiv

  30. [38]

    Wavegan: Frequency-aware gan for high-fidelity few-shot image generation, in: European conference on computer vision, Springer

    Yang, M., Wang, Z., Chi, Z., Feng, W., 2022. Wavegan: Frequency-aware gan for high-fidelity few-shot image generation, in: European conference on computer vision, Springer. pp. 1–17

    2022

  31. [39]

    Time-series generative adversarial networks, in: Advances in Neural Information Processing Systems 32 (NeurIPS 2019), pp

    Yoon, J., Jarrett, D., van der Schaar, M., 2019. Time-series generative adversarial networks, in: Advances in Neural Information Processing Systems 32 (NeurIPS 2019), pp. 5509–5519

    2019

  32. [40]

    Nonintrusiveapplianceloadmonitoring:Reviewandoutlook

    Zeifman,M.,Roth,K.,2011. Nonintrusiveapplianceloadmonitoring:Reviewandoutlook. IEEETransactionsonConsumerElectronics57, 76–84. doi:10.1109/TCE.2011.5735484. First Author et al.:Preprint submitted to ElsevierPage 18 of 18

    2011

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.