acceptodds
Under review as a conference paper at ICLR 2027

Dual-PID-Driven Adaptive Latent Factorization of Tensors for NILM Data Representation

Abstract

Non-intrusive load monitoring (NILM) plays an important role in intelligent energy management systems. However, due to device failures, communication interruptions, and sensor instability, NILM data are often highly incomplete, making it difficult to obtain reliable load representations for downstream tasks. Latent factorization of tensors (LFT) provides an effective solution for NILM data representation by preserving intrinsic multidimensional structures and capturing temporal correlations. However, existing stochastic gradient descent (SGD)-based LFT models often suffer from limited optimization efficiency and representation accuracy. To address these challenges, this paper proposes a dual-proportional-integral-derivative (PID)-driven adaptive latent factorization of tensors (DPAL) model. Its main innovations include: (i) a nonlinear PID controller-driven training error regulation mechanism that exploits historical error information to improve optimization efficiency and representation accuracy; (ii) an incremental PID controller-integrated fine-grained learning rate adjustment strategy that enhances convergence speed and optimization stability; and (iii) a fuzzy logic-based adaptive gain tuning mechanism for PID controllers that improves model practicality and adaptability. Moreover, we provide a conditional coordinate-wise convergence and stability analysis under explicit smoothness and boundedness assumptions. Experimental results on three real-world NILM datasets demonstrate that the proposed DPAL model consistently outperforms state-of-the-art methods in terms of representation accuracy and computational efficiency.

Primary area:
optimization
Keywords:
Non-Intrusive Load Monitoring, Latent Factorization of Tensors, Adaptive Optimization, PID Controller, Representation Learning.

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_wQSbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DPAL, a dual-PID-driven adaptive latent factorization method for reconstructing highly incomplete non-intrusive load monitoring (NILM) data. Built upon CP tensor factorization, DPAL introduces nonlinear PID-based reconstruction-error regulation, incremental PID-based learning-rate adjustment, and fuzzy logic-based online controller tuning. The paper also provides conditional coordinate-wise convergence and gradient-boundedness analyses. Experiments on three real-world NILM datasets, under three observation densities, report improvements in RMSE and MAE over seven tensor factorization or completion baselines. Additional ablations examine the contributions of the two PID controllers and fuzzy adaptation.

Critical strengths
  1. DPAL achieves the lowest reported mean RMSE and MAE across all nine experimental settings on three real-world NILM datasets. The results suggest that the proposed optimization strategy may improve tensor reconstruction under severe data incompleteness.

  2. The method addresses optimization challenges in incomplete tensor factorization and integrates error regulation with adaptive learning-rate adjustment. The paper includes ablations of both PID controllers and fuzzy adaptation, providing initial evidence of their complementary empirical effects.

Critical weaknesses
  1. The experiments lack comparisons with Adam, RMSProp, momentum SGD, and existing PID-based optimizers under the same tensor factorization framework. Consequently, it remains unclear whether the reported improvements originate from the proposed dual-PID design rather than standard adaptive optimization techniques.

  2. The learning-rate update in Eq. (12) does not explicitly guarantee positive learning rates, while the nonlinear PID controller in Eq. (8) accumulates historical errors without an explicit anti-windup mechanism in Algorithm 1. These omissions raise concerns about the stability of the proposed optimization procedure.

  3. The convergence analysis primarily concerns coordinate-wise SGD with fixed regularization and diminishing positive step sizes, rather than the complete DPAL algorithm with adaptive PID controllers and dynamically varying regularization. Therefore, the theoretical results do not directly establish the convergence of the proposed method.

  4. The missing-data generation and normalization procedures are inadequately specified. Moreover, the efficiency comparison unfairly contrasts online adaptation with the total cost of manual grid search, while the separately reported RMSE and MAE training times lack a clear explanation.

Questions for authors

See Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Weakness 1: Optimizer Comparisons

Weaknesses1: The experiments lack comparisons with Adam, RMSProp, momentum SGD, and existing PID-based optimizers under the same tensor factorization framework. Consequently, it remains unclear whether the reported improvements originate from the proposed dual-PID design rather than standard adaptive optimization techniques.

Response: We thank the reviewer for highlighting the importance of direct optimizer comparisons. To clarify whether the improvements arise from the proposed dual-PID mechanism rather than conventional adaptive optimization, we have added the following experiments and explanations.

  1. Matched conventional-optimizer controls. We added SGD, Momentum SGD, Adam, and RMSProp using the same rank-20 biased-CP reconstruction structure, transformed inputs, observation masks, retained-entry splits, initialization, test coordinates, and entry-wise updates. Each conventional optimizer is independently selected by validation RMSE using the same candidate-count budget and regularization grid. DPAL retains fuzzy-adaptive regularization, while these controls use selected fixed regularization. The detailed configurations and selection procedures are provided in Appendix A.3.

  2. Direct PID comparisons and adaptation details. We further compared DPAL with full NPIL, which retains nonlinear PID error regulation and PSO-based gain adaptation, and FPS-CP, which explicitly adapts the existing matrix-based fuzzy PID method to three-factor CP learning. These methods are evaluated under the common data protocol. Their implementation adaptations and selection procedures are reported in Appendix A.3. These comparisons complement the conventional-optimizer controls without assuming identical tuning budgets across different methods.

  3. Accuracy evidence and controller contributions. The following table reports RMSE with standard deviations at 5% observation density. DPAL achieves the lowest mean RMSE across all three datasets. Compared with Adam, the strongest tested conventional optimizer, DPAL reduces RMSE by 4.01%, 5.36%, and 2.71% on iAWE, UK-DALE, and MORED, respectively. Furthermore, SLF removes both controllers under the shared component-ablation settings and is distinct from the independently tuned SGD baseline. The SLF/NPL/IPL/DPAL ablations in Appendix B.3 examine the individual contributions of the two feedback loops, while complete comparisons with related PID methods are provided in Appendix B.1.

RMSE comparison at 5% observation density (mean ± SD).

DatasetDPAL (Ours)SGDMomentumAdamRMSPropNPILFPS-CP
iAWE0.0263 ± 0.000050.02770 ± 0.000210.02777 ± 0.000300.02740 ± 0.000240.02766 ± 0.000260.02672 ± 0.000090.02810 ± 0.00024
UK-DALE0.0247 ± 0.000060.02636 ± 0.000240.02638 ± 0.000250.02610 ± 0.000170.02631 ± 0.000150.02502 ± 0.000080.02630 ± 0.00020
MORED0.0323 ± 0.000100.03342 ± 0.000410.03334 ± 0.000440.03320 ± 0.000320.03361 ± 0.000460.03278 ± 0.000160.03470 ± 0.00040
AuthorsOpenReview ↗

Response to Weakness 2: Learning-Rate Bounds and NPID Stability

Weaknesses 2: The learning-rate update in Eq. (12) does not explicitly guarantee positive learning rates, while the nonlinear PID controller in Eq. (8) accumulates historical errors without an explicit anti-windup mechanism in Algorithm 1. These omissions raise concerns about the stability of the proposed optimization procedure.

Response: We thank the reviewer for raising these important concerns regarding optimization stability. We have clarified the implemented learning-rate constraints and controller-state handling in the revised manuscript, and added experimental evidence to evaluate the numerical stability of DPAL.

  1. Positive and bounded learning rates. To ensure that the learning rate remains positive and bounded, DPAL applies a projection after each IPID update:

    ηijkl=clip⁡(ηijkl−1+Δuijkl, 0.001, 0.010).\eta_{ijk}^{l}=\operatorname{clip}\left(\eta_{ijk}^{l-1}+\Delta u_{ijk}^{l},\,0.001,\,0.010\right).

    This operation guarantees that the learning rate remains within [0.001,0.010][0.001,0.010] throughout training. We have explicitly incorporated this projection into Eq. (12) and Algorithm 1. The initialization and controller settings are further detailed in Appendices A.2 and A.4.

  2. NPID integral behavior and numerical stability. The IPID controller employs an integral-separation strategy to regulate learning-rate adjustments. The NPID controller, however, does not explicitly clip its accumulated integral or impose a hard upper bound on the nonlinear amplification cosh⁡(fme)\cosh(f_m e). To examine its practical stability, we recorded the maximum absolute NPID integral, maximum nonlinear amplification, learning-rate boundary-hit frequencies, and numerical-failure frequency across 20 independent runs. As shown below, no numerical failures occurred in any of the tested settings, and the observed integral and amplification values remained finite. These empirical observations characterize the tested training trajectories rather than establish unconditional stability guarantees.

  3. Sensitivity to learning-rate bounds. We further evaluated three learning-rate intervals on UK-DALE at 5% observation density while keeping the other settings unchanged. As shown in the second table, all three DPAL variants achieve lower mean RMSE and MAE than the fixed-learning-rate NPL baseline. The RMSE values remain close across different intervals, although the learning-rate boundary-hit frequencies vary substantially. These results demonstrate that DPAL maintains consistent reconstruction performance under the tested learning-rate constraints. Additional stability statistics are reported in Appendix B.5.

Observed numerical-stability statistics at 5% observation density.

DatasetLR BoundsMax NPID IntegralMax coshLower HitUpper HitFailed Runs
iAWE0.001–0.0103.1341.13468.2%19.9%0/20
UK-DALE0.001–0.0105.0541.12164.6%16.3%0/20
MORED0.001–0.0103.6951.1563.1%0.0%0/20

Note: The NPID integral is not clipped. The reported maxima are empirical observations from the tested runs.

Learning-rate clipping sensitivity on UK-DALE at 5% observation density (mean ± SD).

VariantLR BoundsRMSEMAELower HitUpper Hit
NPL (Fixed LR)—0.02480 ± 0.000100.01830 ± 0.00010——
DPAL-Tight[0.002, 0.008]0.02476 ± 0.000090.01812 ± 0.0000673.5%20.1%
DPAL (Original)[0.001, 0.010]0.02470 ± 0.000060.01800 ± 0.00000764.6%16.3%
DPAL-Wide[0.0005, 0.015]0.02474 ± 0.000080.01808 ± 0.0000545.2%9.7%
AuthorsOpenReview ↗

Response to Weakness 3: Scope of Convergence Analysis

Weaknesses 3: The convergence analysis primarily concerns coordinate-wise SGD with fixed regularization and diminishing positive step sizes, rather than the complete DPAL algorithm with adaptive PID controllers and dynamically varying regularization. Therefore, the theoretical results do not directly establish the convergence of the proposed method.

Response: We thank the reviewer for pointing out the distinction between the theoretical convergence conditions and the complete DPAL algorithm. We agree that the current analysis does not establish convergence of the fully adaptive DPAL procedure. To clarify the scope of our theoretical contributions, we have revised Section 3.5, Appendix C, and the Conclusion to explicitly distinguish the following results and their assumptions.

  1. Conditional coordinate-wise convergence. Appendix C establishes an O(1/l)O(1/l) expected suboptimality rate for the underlying scalar-coordinate SGD update, with all other factor and bias coordinates held fixed. This result requires fixed regularization λ>0\lambda>0, coordinate-wise smoothness and strong convexity, unbiased stochastic gradients with bounded second moments, and a positive diminishing learning-rate schedule. These conditions support convergence of the coordinate-wise optimization problem, but do not establish joint convergence of the nonconvex tensor factorization objective.

  2. Conditional boundedness of NPID-regulated gradients. We separately analyze the NPID-regulated coordinate gradients under bounded factor states, reconstruction errors, controller gains, nonlinear amplification, and accumulated integral states. Under these assumptions, the regulated gradients remain bounded. However, this result does not establish that the PID-regulated update is an unbiased stochastic gradient of the original objective. Therefore, the coordinate-SGD convergence rate cannot be directly extended to the complete PID-controlled algorithm.

  3. Relationship to the implemented DPAL algorithm. In the actual implementation, learning-rate projection ensures positive and bounded IPID learning rates, but does not guarantee that they diminish over iterations. Furthermore, fuzzy inference dynamically adjusts regularization, while the accumulated NPID integral is not explicitly bounded. Consequently, the complete DPAL algorithm does not necessarily satisfy the assumptions of the coordinate-wise convergence analysis. We have explicitly clarified these limitations in the revised manuscript. The theoretical results are presented as conditional convergence and gradient-boundedness analyses, rather than a convergence guarantee for the complete adaptive DPAL algorithm.

AuthorsOpenReview ↗

Response to Weakness 4: Data Protocol, Normalization, and Efficiency Comparisons

Weaknesses 4: The missing-data generation and normalization procedures are inadequately specified. Moreover, the efficiency comparison unfairly contrasts online adaptation with the total cost of manual grid search, while the separately reported RMSE and MAE training times lack a clear explanation.

Response: We thank the reviewer for these constructive comments regarding experimental reproducibility and computational efficiency. We have clarified the data preprocessing and missingness protocols, distinguished per-fit training time from cumulative hyperparameter-search cost, and explained the metric-specific timing measurements in the revised manuscript.

  1. Data preprocessing and normalization. All compared methods use the same pointwise logarithmic transformation, x′=log⁡(1+x)x'=\log(1+x). The tensors are organized along three axes: sampling time, measurement channel, and monitoring day. Appendix A.1 now provides the data sources, selected houses, recording periods, tensor dimensions, preprocessing procedures, and experimental settings to ensure reproducibility.

  2. Missing-data generation and evaluation protocol. Tube-wise sampling first constructs the retained observation pool Ωobs\Omega_{\mathrm{obs}} at the specified observation density. Its entries are subsequently divided into training, validation, and test sets using an 8:1:1 ratio. The main experiments evaluate randomly held-out entries within this pool, rather than the entirely removed tubes. To further evaluate structured missingness, we conducted separate experiments on UK-DALE using measurement-axis tubes X[t,:,d]\mathcal{X}[t,:,d] and contiguous all-channel gaps X[s:s+L,:,d]\mathcal{X}[s:s+L,:,d], where L∈{600,3600}L\in\{600,3600\}. These test regions are completely excluded from training and validation, and reconstruction errors are calculated only at test coordinates with available ground truth. The results below demonstrate that DPAL achieves lower RMSE and MAE than NPIL across all four evaluation settings. Detailed protocols are provided in Appendices A.1 and B.2.

  3. Clarification of training-time measurements. Validation data determine model selection, early stopping, and checkpoint selection. The separately reported RMSE Time and MAE Time represent the cumulative training times from initialization to the checkpoints achieving the best corresponding validation metrics. These times may differ because the optimal RMSE and MAE checkpoints can occur at different iterations. We additionally report TτT_\tau, defined as the first time a method reaches the selected SGD reference's best validation-RMSE threshold on the same data partition. This provides a common-error criterion for comparing optimization efficiency, independent of each method's final accuracy.

  4. Separation of fitting and tuning costs. We have revised the efficiency analysis to distinguish individual model-fitting time from cumulative hyperparameter-search time. DPAL-M requires repeated fixed-configuration training runs, whereas DPAL employs a predefined fuzzy-rule base shared across all datasets without additional dataset-specific rule-calibration searches. The online fuzzy interpolation overhead is included in DPAL's training time. As shown in the second table, DPAL-M can achieve slightly shorter individual fitting times, while its cumulative search cost is substantially higher. We therefore attribute the overall cost difference to repeated configuration searches rather than claiming uniform per-fit acceleration. The computational timing comparison does not include the human effort involved in designing the predefined fuzzy rules. Complete timing details are provided in Appendix B.6.

Main-protocol and structured-outage comparisons on UK-DALE at 5% observation density (mean ± SD).

Missing PatternDPAL RMSENPIL RMSEDPAL MAENPIL MAETest Entries
Main protocol (random test entries)0.02470 ± 0.000060.02502 ± 0.000080.01800 ± 0.0000070.01825 ± 0.0001036,288
Measurement tube0.02682 ± 0.000160.02703 ± 0.000190.01884 ± 0.000120.01923 ± 0.0001536,288
600-point gap0.02778 ± 0.000230.02855 ± 0.000280.02183 ± 0.000170.02226 ± 0.0002137,296
3600-point gap0.02610 ± 0.000310.02786 ± 0.000370.02099 ± 0.000240.02206 ± 0.0002942,666

Per-fit checkpoint time, cumulative search cost, and common-error time at 5% observation density (seconds).

DatasetDPAL RMSE TimeDPAL-M RMSE TimeDPAL-M Search TotalDPAL TτT_\tauDPAL-M TτT_\tau
iAWE2.992.78232.870.750.82
UK-DALE2.882.39251.830.620.59
MORED0.360.4119.760.040.06
Reviewer_gEUsOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DPAL to recover missing electrical measurements through tensor factorization. It uses two PID controllers to adjust the reconstruction error and learning rate, with fuzzy rules adapting the controller gains and regularization. Experiments on three NILM datasets report lower recovery errors across nine sparsity settings, and ablations examine the two controllers and the fuzzy adaptation.

Critical strengths

The reported recovery results are consistently good: DPAL achieves the lowest RMSE and MAE in all nine settings. The four-way SLF/NPL/IPL/DPAL ablation is useful, showing that error regulation contributes most of the accuracy improvement while the combined method often reduces iterations. The paper also explicitly limits its theory and acknowledges the runtime tradeoff.

Critical weaknesses
  1. A comparison with the closely related NPIL method is missing. Wang et al. (IJCNN 2025) [1] already study nonlinear PID error regulation and gain adaptation for tensor-based NILM recovery, as described in their available author preprint. DPAL changes the nonlinear regulation and adds a learning-rate controller and fuzzy adaptation, but its NPL ablation is not a comparison with the full NPIL method. Comparing accuracy and training/tuning costs on the same splits would help establish what these additions offer.

  2. I could not work out the actual optimizer settings from the paper. Table 2 gives symbols and ordering rules, but not the numerical gain levels or error thresholds; please provide these, the data scaling, and the initial settings. Eq. (12) can also make the learning rate negative, so any bounds used in practice should be stated. More importantly, how were the fuzzy rules chosen, and were they shared across datasets? Since Table 4 includes grid search in DPAL-M's cost, the tuning-cost comparison also needs the search budget and the cost of choosing DPAL's rules.

  3. It is unclear whether the test set includes the continuous gaps the method is meant to recover. Appendix B.1 introduces tube-wise missingness, but Section 4.1 then randomly splits the observed entries into training, validation, and test sets. Read literally, the removed tubes would not be scored. Please explain how the masks and test set are constructed. If the removed intervals are not currently evaluated, results on those intervals are needed to support the claim about continuous data loss.

[1] Wang Y, Xie T, Wu H. Non-Intrusive Load Monitoring Missing Data Recovery via Nonlinear PID-Incorporated Latent Factorization of Tensors[C]//2025 International Joint Conference on Neural Networks (IJCNN). IEEE, 2025: 1-8.

Questions for authors

Please see the weaknesses above.

5 comments
AuthorsOpenReview ↗

Response to Weakness 1: Novelty and Direct NPIL Comparison

Weakness 1: A comparison with the closely related NPIL method is missing. Wang et al. (IJCNN 2025) [1] already study nonlinear PID error regulation and gain adaptation for tensor-based NILM recovery, as described in their available author preprint. DPAL changes the nonlinear regulation and adds a learning-rate controller and fuzzy adaptation, but its NPL ablation is not a comparison with the full NPIL method. Comparing accuracy and training/tuning costs on the same splits would help establish what these additions offer.

Response: We thank the reviewer for highlighting the importance of comparing DPAL with the closely related NPIL method. We have added direct comparisons with full NPIL under the same data partitions and evaluation protocol, and clarified the methodological differences and computational costs in the revised manuscript.

  1. Direct comparison with full NPIL. We have evaluated the complete NPIL method, retaining its nonlinear PID error regulation and PSO-based gain adaptation during training. We clarify that NPL is an ablation of DPAL that retains only the NPID controller and therefore cannot replace a comparison with full NPIL. The baseline implementation is described in Appendix A.3, and complete RMSE and MAE comparisons across all nine experimental settings are provided in Appendix B.1.

  2. Distinction from NPIL. Although both methods employ nonlinear PID-based error regulation, their learning mechanisms differ. NPIL uses separate sech⁡\operatorname{sech} and exponential mappings for its P/I/D gains, with the integral gain applied to the accumulated residual. DPAL instead applies cosh⁡(fme)\cosh(f_m e) to individual historical residuals before accumulation and introduces an additional entry-specific IPID controller to regulate the learning step. Furthermore, shared fuzzy inference adaptively coordinates both PID controllers, nonlinear amplification, and regularization during factor and bias updates. The SLF/NPL/IPL/DPAL ablations in Appendix B.3 further examine the contributions of these components.

  3. Accuracy and computational costs. As shown in the following tables, DPAL achieves lower mean RMSE and MAE than full NPIL across all nine settings. At 5% observation density, the RMSE reductions are 1.57%, 1.28%, and 1.46% on iAWE, UK-DALE, and MORED, respectively. We also report the measured training times, including the cost of NPIL's in-training PSO adaptation. To distinguish optimization efficiency from final reconstruction accuracy, we report both the elapsed time to each method's best validation-RMSE checkpoint and TτT_\tau, the time required to reach a common validation-RMSE threshold defined by the selected SGD reference. DPAL reaches this common target sooner on all three datasets, although NPIL reaches its own best-RMSE checkpoint sooner on MORED. These measurements are reported in Appendix B.6; a separate offline NPIL tuning-search total is not included.

Direct comparison between DPAL and full NPIL across all nine settings (mean ± SD).

DatasetDPAL RMSENPIL RMSEDPAL MAENPIL MAE
iAWE 5%0.0263 ± 0.000050.02672 ± 0.000090.0187 ± 0.000060.01898 ± 0.00010
iAWE 10%0.0258 ± 0.000030.02616 ± 0.000060.0183 ± 0.000050.01856 ± 0.00021
iAWE 15%0.0256 ± 0.000080.02598 ± 0.000040.0181 ± 0.000040.01838 ± 0.00015
UK-DALE 5%0.0247 ± 0.000060.02502 ± 0.000080.0180 ± 0.0000070.01825 ± 0.00010
UK-DALE 10%0.0246 ± 0.000010.02497 ± 0.000180.0177 ± 0.000010.01796 ± 0.00025
UK-DALE 15%0.0244 ± 0.0000070.02472 ± 0.000130.0173 ± 0.0000030.01755 ± 0.00010
MORED 5%0.0323 ± 0.000100.03278 ± 0.000160.0252 ± 0.000200.02556 ± 0.00018
MORED 10%0.03173 ± 0.000080.03225 ± 0.000300.0248 ± 0.000100.02515 ± 0.00029
MORED 15%0.03155 ± 0.000070.03206 ± 0.000100.0247 ± 0.000050.02506 ± 0.00020

Training-time comparison between DPAL and full NPIL at 5% observation density (mean ± SD).

DatasetMethodRMSERMSE Time (s)TτT_\tau (s)
iAWEDPAL0.02630 ± 0.000052.99 ± 0.730.75 ± 0.07
iAWENPIL0.02672 ± 0.000094.15 ± 0.212.30 ± 0.16
UK-DALEDPAL0.02470 ± 0.000062.88 ± 0.740.62 ± 0.06
UK-DALENPIL0.02502 ± 0.000083.24 ± 0.181.95 ± 0.12
MOREDDPAL0.03230 ± 0.000100.36 ± 0.060.04 ± 0.01
MOREDNPIL0.03278 ± 0.000160.32 ± 0.030.17 ± 0.03
AuthorsOpenReview ↗

Response to Weakness 2: Numerical Settings, Fuzzy Rules, and Tuning Costs

Weakness 2: I could not work out the actual optimizer settings from the paper. Table 2 gives symbols and ordering rules, but not the numerical gain levels or error thresholds; please provide these, the data scaling, and the initial settings. Eq. (12) can also make the learning rate negative, so any bounds used in practice should be stated. More importantly, how were the fuzzy rules chosen, and were they shared across datasets? Since Table 4 includes grid search in DPAL-M's cost, the tuning-cost comparison also needs the search budget and the cost of choosing DPAL's rules.

Response: We thank the reviewer for emphasizing numerical reproducibility and fair tuning-cost comparisons. We have specified the complete optimizer settings, clarified the fuzzy-rule design and learning-rate safeguards, and separated individual fitting time from cumulative search cost.

  1. Numerical settings and initialization. Appendix A.2 now provides all five error anchors and 40 parameter levels, summarized below. Factors and biases are initialized from U(0,0.05)\mathcal{U}(0,0.05), controller histories from zero, and the initial learning rate is 0.002. The IPID integral-separation threshold is σ=0.05\sigma=0.05. All methods use x′=log⁡(1+x)x'=\log(1+x). To ensure positive, bounded learning rates, Eq. (12) and Algorithm 1 explicitly apply ηijkl=clip⁡(ηijkl−1+Δuijkl,0.001,0.010)\eta_{ijk}^{l}=\operatorname{clip}(\eta_{ijk}^{l-1}+\Delta u_{ijk}^{l},0.001,0.010).

  2. Fuzzy-rule design and selection. The rules are prescribed before training and shared unchanged across iAWE, UK-DALE, and MORED, without dataset-specific rule searches. Larger errors receive stronger proportional action, while smaller errors receive greater integral/derivative weighting and regularization. Adjacent fuzzy memberships provide continuous interpolation. We clarify that these numerical levels are predefined design choices, and their human design effort is not quantified by the computational timings.

  3. Sensitivity analysis. We added one-group perturbation experiments on UK-DALE at 5% density (Appendix B.4). Scaling controller gains and regularization by 0.8/1.2 and error anchors by 0.5/2.0 increases mean RMSE by at most 1.62%. Reducing regularization slightly improves MAE from 0.01800 to 0.01798. Three adaptive learning-rate intervals yield RMSE between 0.02470 and 0.02476 (Appendix B.5). These results indicate limited sensitivity under the tested perturbations without establishing optimality of the prescribed settings.

  4. Search budget and computational costs. Appendix A.3 specifies DPAL-M's coordinate search: eight parameters, three candidate values per parameter, and at most five sweeps, yielding up to 81 distinct fits. The second table separates selected-fit checkpoint time from cumulative search cost. DPAL-M is slightly faster per fit on iAWE and UK-DALE, whereas its cumulative search cost is substantially higher across all three datasets. DPAL uses one shared rule base without additional calibration searches, and its online interpolation cost is included in training time. Thus, the search-cost difference reflects repeated fixed-configuration fits rather than uniform per-fit acceleration (Appendix B.6).

Numerical fuzzy-rule levels shared across all datasets.

ParameterLevel 1Level 2Level 3Level 4Level 5
Error anchors0.00010.0010.010.11
KpK_p5.65.86.06.26.4
KiK_i0.010.0080.0060.0040.002
KdK_d0.0050.0040.0030.0020.001
fmf_m1.00.90.80.70.6
CpC_p0.0010.0020.0030.0040.005
CiC_i0.010.00950.0090.0080.007
CdC_d0.120.110.100.090.08
λ\lambda0.030.0250.020.0150.01

Per-fit checkpoint time, cumulative search cost, and common-error time at 5% observation density (seconds).

DatasetDPAL RMSE TimeDPAL-M RMSE TimeDPAL-M Search TotalDPAL TτT_\tauDPAL-M TτT_\tau
iAWE2.992.78232.870.750.82
UK-DALE2.882.39251.830.620.59
MORED0.360.4119.760.040.06
AuthorsOpenReview ↗

Response to Weakness 3: Evaluation of Tube-Wise and Continuous Missingness

Weakness 3: It is unclear whether the test set includes the continuous gaps the method is meant to recover. Appendix B.1 introduces tube-wise missingness, but Section 4.1 then randomly splits the observed entries into training, validation, and test sets. Read literally, the removed tubes would not be scored. Please explain how the masks and test set are constructed. If the removed intervals are not currently evaluated, results on those intervals are needed to support the claim about continuous data loss.

Response: We thank the reviewer for identifying the ambiguity in our missing-data evaluation protocol. We agree that the original random-entry evaluation does not directly measure reconstruction accuracy within entirely removed tubes or continuous gaps. To address this concern, we have clarified the data partitioning procedure and added separate structured-missingness experiments in the revised manuscript.

  1. Clarification of the original evaluation protocol. Tube-wise sampling first constructs a retained observation pool at each specified density (5%, 10%, and 15%). The retained entries are then randomly divided into training, validation, and test sets using an 8:1:1 ratio. Therefore, the original experiments evaluate randomly held-out entries within the retained pool, rather than the complete tubes removed during sampling. We have explicitly clarified this distinction in Section 4.1 and Appendix A.1.

  2. Additional tube-wise and continuous-gap experiments. To directly evaluate structured missingness, we conducted additional experiments on UK-DALE at 5% observation density. The tensor is organized along sampling-time, measurement-channel, and monitoring-day dimensions. We considered three structured missingness patterns: (i) measurement tubes containing all channels at a selected sampling time and day; (ii) continuous gaps spanning 600 sampling points; and (iii) continuous gaps spanning 3600 sampling points. Both gap types remove all measurement channels throughout the selected interval. The designated test regions are completely excluded from training and validation, and RMSE/MAE are calculated only over test coordinates with measured ground truth. Separate validation coordinates determine checkpoint selection and stopping. The detailed construction and evaluation procedures are provided in Appendix A.1.

  3. Experimental evidence. The following table reports the original random-entry control and the three additional structured-missingness experiments. DPAL consistently achieves lower mean RMSE and MAE than full NPIL across all four settings. For the 600-point and 3600-point gaps, DPAL reduces RMSE by 2.70% and 6.32%, respectively, and MAE by 1.93% and 4.85%. These results demonstrate that the reconstruction improvements extend to continuous missing intervals excluded from the training data. The complete results, including standard deviations and test-entry counts, are reported in Appendix B.2.

Main-protocol control and structured-missingness evaluation on UK-DALE at 5% observation density (mean ± SD).

Missing PatternDPAL RMSENPIL RMSEDPAL MAENPIL MAETest Entries
Main protocol (random test entries)0.02470 ± 0.000060.02502 ± 0.000080.01800 ± 0.0000070.01825 ± 0.0001036,288
Measurement tube0.02682 ± 0.000160.02703 ± 0.000190.01884 ± 0.000120.01923 ± 0.0001536,288
600-point gap0.02778 ± 0.000230.02855 ± 0.000280.02183 ± 0.000170.02226 ± 0.0002137,296
3600-point gap0.02610 ± 0.000310.02786 ± 0.000370.02099 ± 0.000240.02206 ± 0.0002942,666
Reviewer_K1esOpenReview ↗
Rating
4(+2)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DPAL combines biased CP tensor factorization with nonlinear PID error regulation, incremental PID learning-rate adjustment, and fuzzy gain/regularization adaptation. Experiments evaluate missing-value reconstruction on three NILM datasets at three observation densities, alongside controller and adaptation ablations.

Critical strengths
  1. DPAL achieves the lowest reported mean RMSE and MAE in all nine settings, with standard deviations over 20 runs. The datasets cover different sampling rates and severe missingness.

  2. The controller ablations examine accuracy, iterations, and runtime. The paper clearly limits its convergence analysis to conditional coordinate SGD and acknowledges that fewer iterations do not always yield faster training.

Critical weaknesses
  1. Novelty is insufficiently established against close precedents. NPIL [1] already applies nonlinear PID error regulation to tensor factorization for NILM, while FPS [2] uses fuzzy PID gain/hyperparameter adaptation for factor analysis. Neither is discussed or benchmarked, and the incremental PID scheduler is explicitly borrowed from Wang and Zhang (2024). Explain the distinct contribution and compare directly; the NPL ablation does not replace an NPIL comparison.

  2. The baselines include a 2025 method, but omit Adam/momentum on the same biased CP objective and the cited fast CP-completion approach of Ghadiri et al. (2025) [3]. Comparisons across different factorization models cannot isolate the optimizer's benefit. Match inputs, objectives, initialization, and tuning budgets for optimizer comparisons, and report baseline adaptation details.

  3. Appendix A.1 leaves all five error reference points and 40 parameter levels symbolic; initial learning rates, the integral-separation threshold, normalization, and dataset houses/dates are also unspecified. Online interpolation still depends on these design choices, so the claim of avoiding manual tuning needs qualification. Provide numerical settings, selection procedures, and sensitivity results.

  4. Equation (12) increments the learning rate using signed residuals without ensuring positivity or imposing an upper bound. Algorithm 1 also lacks explicit safeguards for the NPID integral and cosh amplification. Appendix C correctly restricts its rate to fixed-regularization coordinate SGD, leaving the actual dual-loop trajectory uncovered. Specify implemented safeguards and report learning-rate/integral trajectories and instability frequency.

  5. Appendix B.1 describes tube-wise and continuous missingness, whereas Section 4.1 randomly splits observed entries 8:1:1. It is unclear whether the test set contains held-out tubes or only scattered entries. Specify mask axes and gap lengths, evaluate block outages separately, and identify which split controls stopping. Reconstruction errors alone also do not establish downstream NILM utility.

  6. The cost advantage in Table 4 includes avoided manual grid search (459.56 s versus 5.94 s), leaving per-fit acceleration unclear. Separate tuning and fitting costs, include fuzzy-rule selection costs, and compare time to a common validation error. Report uncertainty for the dual-controller ablations, where some accuracy differences are small.

[1] Wang, Y., Xie, T., and Wu, H. Latent Tensor Factorization with Nonlinear PID Control for Missing Data Recovery in Non-Intrusive Load Monitoring. arXiv:2504.13483, 2025.

[2] Li, J., and Yuan, Y. Fast Latent Factor Analysis via a Fuzzy PID-Incorporated Stochastic Gradient Descent Algorithm. arXiv:2303.03941, 2023.

[3] Ghadiri, M., Fahrbach, M., Kook, Y., and Jadbabaie, A. Fast Tensor Completion via Approximate Richardson Iteration. ICML, PMLR 267:19248-19265, 2025.

Questions for authors
  1. What distinguishes DPAL from NPIL and FPS, and do its gains persist against these methods and Adam/momentum under matched optimization and tuning budgets?

  2. What numerical fuzzy rules and initial settings were used, how were they selected, and what are the separate tuning and per-fit costs?

  3. How are positive, bounded learning rates and bounded NPID integral states enforced? Which conditions in Appendix C hold for the implemented algorithm?

  4. How are tube/block masks constructed and assigned to the test set, which split controls stopping, and does reconstruction improve a downstream NILM task?

8 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1: Novelty and Related PID Methods

Weakness 1: Novelty is insufficiently established against close precedents. NPIL [1] already applies nonlinear PID error regulation to tensor factorization for NILM, while FPS [2] uses fuzzy PID gain/hyperparameter adaptation for factor analysis. Neither is discussed or benchmarked, and the incremental PID scheduler is explicitly borrowed from Wang and Zhang (2024). Explain the distinct contribution and compare directly; the NPL ablation does not replace an NPIL comparison.

Question 1: What distinguishes DPAL from NPIL and FPS, and do its gains persist against these methods and Adam/momentum under matched optimization and tuning budgets?

Response: We thank the reviewer for highlighting the need to establish the contribution of DPAL relative to existing PID-based methods. We have expanded the related-work discussion and added direct experimental comparisons to clarify the methodological differences and performance improvements.

  1. Dual-feedback contribution. The main contribution of DPAL is a coupled optimization mechanism that combines nonlinear PID (NPID) residual regulation with an entry-specific incremental PID (IPID) learning-rate controller for biased-CP factorization. Shared fuzzy inference jointly adapts both controllers, nonlinear amplification, and regularization during factor and bias updates. We explicitly acknowledge that the incremental PID formulation originates from Wang and Zhang (2024). Our contribution lies in its integration with nonlinear error regulation and coordinated fuzzy adaptation within the dual-feedback learning framework.

  2. Distinction from NPIL and FPS. Full NPIL combines nonlinear PID error regulation with PSO-based gain adaptation. DPAL additionally regulates entry-specific learning rates and applies cosh⁡(fme)\cosh(f_m e) to individual historical residuals before accumulation, whereas NPIL employs separate sech⁡\operatorname{sech} and exponential P/I/D gain mappings. FPS focuses on matrix factorization; we adapt it to three-factor CP updates as FPS-CP for direct evaluation. The revised Introduction now discusses these differences and acknowledges the corresponding prior work.

  3. Direct experimental evidence. We compared DPAL with full NPIL, including its PSO gain adaptation, and FPS-CP under the common evaluation protocol. As shown below, DPAL achieves lower mean RMSE and MAE than both methods across all nine settings. The NPL ablation is evaluated separately and is not used as a substitute for full NPIL. Additional comparisons with validation-tuned Adam, Momentum SGD, and other conventional optimizers are provided in our response to CW2, while controller ablations and timing results are reported in Appendices B.3 and B.6. Complete PID-method comparisons are provided in Appendix B.1.

Direct comparison with related PID methods across all nine settings (mean ± SD).

DatasetDPAL RMSENPIL RMSEFPS-CP RMSEDPAL MAENPIL MAEFPS-CP MAE
iAWE 5%0.0263 ± 0.000050.02672 ± 0.000090.02810 ± 0.000240.0187 ± 0.000060.01898 ± 0.000100.01965 ± 0.00017
iAWE 10%0.0258 ± 0.000030.02616 ± 0.000060.02710 ± 0.000180.0183 ± 0.000050.01856 ± 0.000210.01920 ± 0.00013
iAWE 15%0.0256 ± 0.000080.02598 ± 0.000040.02670 ± 0.000160.0181 ± 0.000040.01838 ± 0.000150.01895 ± 0.00012
UK-DALE 5%0.0247 ± 0.000060.02502 ± 0.000080.02630 ± 0.000200.0180 ± 0.0000070.01825 ± 0.000100.01910 ± 0.00015
UK-DALE 10%0.0246 ± 0.000010.02497 ± 0.000180.02590 ± 0.000180.0177 ± 0.000010.01796 ± 0.000250.01880 ± 0.00013
UK-DALE 15%0.0244 ± 0.0000070.02472 ± 0.000130.02560 ± 0.000160.0173 ± 0.0000030.01755 ± 0.000100.01848 ± 0.00011
MORED 5%0.0323 ± 0.000100.03278 ± 0.000160.03470 ± 0.000400.0252 ± 0.000200.02556 ± 0.000180.02680 ± 0.00026
MORED 10%0.03173 ± 0.000080.03225 ± 0.000300.03390 ± 0.000340.0248 ± 0.000100.02515 ± 0.000290.02620 ± 0.00023
MORED 15%0.03155 ± 0.000070.03206 ± 0.000100.03340 ± 0.000290.0247 ± 0.000050.02506 ± 0.000200.02570 ± 0.00020
AuthorsOpenReview ↗

Response to Weakness 2 and Question 1: Optimizer Comparisons and Fairness

Weakness 2: The baselines include a 2025 method, but omit Adam/momentum on the same biased CP objective and the cited fast CP-completion approach of Ghadiri et al. (2025) [3]. Comparisons across different factorization models cannot isolate the optimizer's benefit. Match inputs, objectives, initialization, and tuning budgets for optimizer comparisons, and report baseline adaptation details.

Question 1: What distinguishes DPAL from NPIL and FPS, and do its gains persist against these methods and Adam/momentum under matched optimization and tuning budgets?

Response: We thank the reviewer for emphasizing the importance of matched optimizer comparisons. We have added conventional adaptive optimizers and FastTC, clarified their implementation and tuning procedures, and evaluated DPAL under a common reconstruction protocol.

  1. Matched conventional-optimizer controls. We added SGD, Momentum SGD, Adam, and RMSProp using the same rank-20 biased-CP reconstruction structure, transformed inputs, observation masks, train/validation/test partitions, initialization procedure, and entry-wise training protocol as DPAL. The four conventional optimizers use the same candidate-count budget and validation-based selection procedure, with their hyperparameter grids detailed in Appendix A.3. DPAL employs fuzzy-adaptive regularization, whereas these controls use validation-selected fixed regularization. We also distinguish the controller-off SLF ablation from the independently tuned SGD baseline.

  2. FastTC and comparison scope. We additionally evaluated FastTC using its released non-accelerated Richardson solver. Appendix A.3 specifies its adaptation, solver limits, and hyperparameter-search procedure. The conventional optimizer controls isolate optimization performance within the shared biased-CP structure, whereas FastTC is evaluated as a separate tensor-completion method. The related PID methods, including full NPIL and FPS-CP, are compared directly in our response to CW1 and Appendix B.1.

  3. Experimental results. As shown below, DPAL achieves the lowest mean RMSE among all compared methods on the three datasets at 5% observation density. Compared with Adam, the strongest conventional optimizer, DPAL reduces RMSE by 4.01%, 5.36%, and 2.71% on iAWE, UK-DALE, and MORED, respectively. These results demonstrate that DPAL's accuracy improvements persist against independently tuned conventional optimizers under the matched reconstruction protocol. Complete baseline settings are provided in Appendix A.3.

RMSE comparison with conventional optimizers and FastTC at 5% observation density (mean ± SD).

DatasetDPAL (Ours)SGDMomentum SGDAdamRMSPropFastTC
iAWE0.0263 ± 0.000050.02770 ± 0.000210.02777 ± 0.000300.02740 ± 0.000240.02766 ± 0.000260.03450 ± 0.00110
UK-DALE0.0247 ± 0.000060.02636 ± 0.000240.02638 ± 0.000250.02610 ± 0.000170.02631 ± 0.000150.03200 ± 0.00125
MORED0.0323 ± 0.000100.03342 ± 0.000410.03334 ± 0.000440.03320 ± 0.000320.03361 ± 0.000460.04500 ± 0.00180
AuthorsOpenReview ↗

Response to Weakness 3 and Question 2: Numerical Rules and Reproducibility

Weakness 3: Appendix A.1 leaves all five error reference points and 40 parameter levels symbolic; initial learning rates, the integral-separation threshold, normalization, and dataset houses/dates are also unspecified. Online interpolation still depends on these design choices, so the claim of avoiding manual tuning needs qualification. Provide numerical settings, selection procedures, and sensitivity results.

Question 2: What numerical fuzzy rules and initial settings were used, how were they selected, and what are the separate tuning and per-fit costs?

Response: We thank the reviewer for identifying the missing implementation details. We have provided the complete numerical rules, initialization settings, sensitivity analysis, and separate fitting/search costs in the revised manuscript.

  1. Numerical settings and rule selection. Appendix A.2 specifies all five error anchors and 40 parameter levels, listed below. The rules are prescribed before training and shared unchanged across iAWE, UK-DALE, and MORED, without dataset-specific rule searches. Larger errors receive stronger proportional action, while smaller errors receive greater integral/derivative weighting and regularization. Adjacent memberships interpolate the parameters. Factors and biases are initialized from U(0,0.05)\mathcal{U}(0,0.05), controller histories from zero, and the initial learning rate is 0.002. The IPID integral-separation threshold is σ=0.05\sigma=0.05, with learning-rate bounds [0.001,0.010][0.001,0.010].

  2. Data preprocessing and sensitivity. All methods use x′=log⁡(1+x)x'=\log(1+x). Appendix A.1 specifies iAWE Building 1 (28 days), UK-DALE House 1 (28 days), and MORED Building 1 (22 days), including sampling rates and tensor dimensions. On UK-DALE at 5% density, we varied one parameter group at a time while fixing the remaining settings. The largest mean-RMSE increase is 1.62%. Reducing regularization slightly improves MAE from 0.01800 to 0.01798. These results support limited sensitivity under the tested perturbations without establishing optimality of the original rules (Appendix B.4).

  3. Tuning and fitting costs. DPAL-M uses coordinate search over eight parameters, three candidate values per parameter, and at most five sweeps (up to 81 distinct fits). At 5% density, DPAL/DPAL-M selected-fit checkpoint times are 2.99/2.78 s (iAWE), 2.88/2.39 s (UK-DALE), and 0.36/0.41 s (MORED). DPAL-M cumulative search costs are 232.87, 251.83, and 19.76 s, respectively. DPAL uses its predefined shared rules without additional dataset-specific calibration searches; online fuzzy interpolation is included in training time. We clarify that avoiding repeated configuration searches does not eliminate human rule-design effort, which is not included in these timings (Appendices A.3 and B.6).

Numerical fuzzy-rule levels shared across all datasets.

ParameterLevel 1Level 2Level 3Level 4Level 5
Error anchors0.00010.0010.010.11
KpK_p5.65.86.06.26.4
KiK_i0.010.0080.0060.0040.002
KdK_d0.0050.0040.0030.0020.001
fmf_m1.00.90.80.70.6
CpC_p0.0010.0020.0030.0040.005
CiC_i0.010.00950.0090.0080.007
CdC_d0.120.110.100.090.08
λ\lambda0.030.0250.020.0150.01

Fuzzy-rule sensitivity on UK-DALE at 5% observation density (mean ± SD).

Parameter GroupScalingRMSEMAE
Original DPAL1.00.02470 ± 0.000060.01800 ± 0.000007
NPID gains (Kp,Ki,Kd)(K_p,K_i,K_d)0.80.02510 ± 0.000120.01834 ± 0.00012
NPID gains (Kp,Ki,Kd)(K_p,K_i,K_d)1.20.02502 ± 0.000140.01827 ± 0.00008
IPID gains (Cp,Ci,Cd)(C_p,C_i,C_d)0.80.02493 ± 0.000100.01816 ± 0.00007
IPID gains (Cp,Ci,Cd)(C_p,C_i,C_d)1.20.02489 ± 0.000100.01814 ± 0.00007
Error anchors Δ\Delta0.50.02486 ± 0.000080.01809 ± 0.00005
Error anchors Δ\Delta2.00.02505 ± 0.000110.01826 ± 0.00009
Regularization λ\lambda0.80.02482 ± 0.000090.01798 ± 0.00005
Regularization λ\lambda1.20.02490 ± 0.000090.01812 ± 0.00006

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.