Source-linked AI summary
Multi-Appliance Non-Intrusive Load Monitoring via Label-Preserving Aggregate Recomposition and Prediction Consistency
Jiangfeng Liu, Yanfang Fan
TL;DR
NILM models trained on source households may lose accuracy on unseen households, where residual-background correlations can differ. The paper combines label-preserving aggregate recomposition with prediction consistency during training. On REDD, UK-DALE, and REFIT, this method lowers macro MAE relative to single-window training.
Problem
NILM models trained on source households may degrade on unseen households, where reliance on residual-background correlations may become less accurate.
Source: A. Background and Motivation · p. 1Method
The method recomposes aggregate windows by changing only residual background while preserving target labels, then selectively constrains paired power predictions during training.
Source: 1) Label-Preserving Aggregate… · p. 2, C. Prediction Consistency · p. 5Results
Macro MAE falls from 14.75 to 13.14 W on REDD, from 8.88 to 8.51 W on UK-DALE, and from 15.83 to 14.55 W on REFIT versus single-window training.
Source: §V · p. 8Takeaways & Limitations
The training method adds no inference-time module or parameter.
Source: §V · p. 8Method-comparison evidence covers three data sets, one unseen household per data set, and one fixed consistency weight.
Source: §V · p. 8
Abstract
from arXiv · showhide
Non-intrusive load monitoring (NILM) estimates appliance power sequences from aggregate power, but models trained on source households commonly lose accuracy in unseen households. Aggregate power also contains loads from other appliances and measurement error, so predictions may depend on the residual background that co-occurs with source-household targets. Time-aligned submetered measurements and the additive decomposition of aggregate power expose a relation unused by window-wise supervision: an aggregate window can be recomposed by replacing only its residual background while preserving all modeled target-appliance power sequences pointwise. We combine label-preserving aggregate recomposition with prediction consistency. Both windows receive complete power and operating-state supervision. For each appliance, disagreement between the two power predictions is penalized only when both satisfy a fixed reliability criterion and only to the extent that it exceeds a fixed margin. The proposed method is implemented using a multi-appliance architecture with two-stage shared-to-specific mixture-of-experts routing. On REDD, UK-DALE, and REFIT, the proposed method lowers appliance-averaged mean absolute error relative to single-window training from 14.75 to 13.14 W, from 8.88 to 8.51 W, and from 15.83 to 14.55 W. Label-preserving aggregate recomposition and prediction consistency are used only during training, and add no inference-time module or parameter.
I. INTRODUCTION
NILM estimates appliance power from aggregate smart-meter readings, but models trained on source households may degrade on unseen households. This paper proposes label-preserving aggregate recomposition with prediction consistency and implements it using FLAME.
A. Background and Motivation
- Unseen-household degradation may reflect differences in target-appliance characteristics, usage habits, and non-target loads across households.The residual background can correlate with target operation in training, and predictions relying on those correlations may become less accurate when they do not hold in other households.Source: A. Background and Motivation · p. 1
B. Related Work and Research Gap
- The reviewed approaches do not provide pairs that differ only in residual background while preserving every target-appliance power sequence pointwise.Conventional window-wise supervision evaluates each window against its targets but does not explicitly encode the relation between paired predictions.Source: B. Related Work and Research Gap · pp. 1–2
C. Contributions of the Present Work
- The proposed method combines label-preserving aggregate recomposition with prediction consistency during training, implemented using the FLAME architecture.FLAME shares aggregate-level context across appliances while forming a distinct representation for each.Source: C. Contributions of the Present Work · p. 2
1) Label-Preserving Aggregate Recomposition and Prediction Consistency:
The method uses the additive decomposition of aggregate power to construct paired windows with identical target labels. It supervises both windows and selectively constrains their per-appliance power predictions.
1) Label-Preserving Aggregate Recomposition and Prediction Consistency:
- Both windows receive complete power and operating-state supervision, while prediction disagreement is penalized only when both predictions meet a reliability criterion and exceed a fixed margin.The penalty acts on gated power outputs, not representations, and does not require identical predictions.Source: 1) Label-Preserving Aggregate… · p. 2
II. PROBLEM STATEMENT
- Residual background is aggregate power not represented by target-appliance sequences, including measurement and preprocessing discrepancies.Time-aligned submetered measurements and additive decomposition make the fixed-target recomposition relation available in this labeled setting.Source: §II · p. 2
- For fixed target-appliance sequences, varying residual background yields multiple aggregate windows with identical targets by construction.With residual background held fixed, substituting a target sequence instead creates a different aggregate whose targets must be recomputed; only the fixed-target direction yields identical labels.Source: §II · p. 2
III. PROPOSED METHOD
The proposed method trains on paired windows and constrains prediction consistency under window-wise supervision. FLAME supplies joint multi-appliance power and operating-state predictions through shared-to-specific expert routing.
A. FLAME Multi-Appliance Architecture
- FLAME first pools appliance-branch responses into a shared representation, then refines it into appliance-specific representations for prediction.The architecture uses a residual temporal-convolutional backbone and two ordered routing stages.Source: A. FLAME Multi-Appliance Architecture · pp. 2–3
- Fig. 1 depicts the backbone, many-to-one aggregation, one-to-many refinement, appliance prediction heads, and routing operator.Its panels also expand the ResTCN backbone and show the routing operator used in the two stages.Source: Figure 1 · p. 3
- Each routed branch mixes candidate expert responses using sample-wise softmax weights and interpolates the mixture with its own gated view.Separate gates determine expert composition and how far the routed output departs from the branch’s own view.Source: A. FLAME Multi-Appliance Architecture · p. 3
1) Cross-Appliance Aggregation:
Cross-appliance aggregation routes a shared representation through a shared expert and appliance-conditioned experts. The updated shared representation alone passes to the next aggregation level.
1) Cross-Appliance Aggregation:
- The aggregation levels precede the final appliance-specific refinement level in FLAME’s routing hierarchy.The shared representation is updated at each aggregation level.Source: 1) Cross-Appliance Aggregation · p. 4
- Each aggregation level routes the shared representation using a shared temporal-convolution response and two heterogeneous expert responses per appliance.The appliance-specific BiGRU and BiMamba experts operate on the same appliance view, while the shared expert operates on the shared view.Source: 1) Cross-Appliance Aggregation · p. 4
- Only the updated shared representation is passed to the next routing level.Source: 1) Cross-Appliance Aggregation · p. 4
2) Appliance-Specific Refinement:
The final routing level gives each appliance its own representation and predicts both power and operating state. Separately, residual-based window construction supports additional training examples while preserving the additive relation.
2) Appliance-Specific Refinement:
- Each appliance uses an independent three-way router to combine the shared response with that appliance’s two expert responses into its own representation.The final routing level refines a single shared representation into appliance-specific representations.Source: 2) Appliance-Specific Refinement · p. 4
- Separate regression and state-classification heads produce gated power from their outputs, with the state targets thresholded from raw appliance power.The gated power is a parameter-free function of the two head outputs.Source: 2) Appliance-Specific Refinement · p. 4
B. Residual-Based Window Construction
- Holding a recorded window’s residual background fixed while substituting target-appliance sequences produces further training windows with targets recomputed from those sequences.The additive relation holds exactly for each constructed window.Source: B. Residual-Based Window Construction · p. 4
1) Training-Window Synthesis:
Training-window synthesis targets sparse appliance activations by retaining a host window’s residual background and substituting an activation segment. Aggregate recomposition instead changes the residual background while preserving the modeled appliance targets.
1) Training-Window Synthesis:
- Sparse target appliances occupy few training samples, so uniformly drawn windows provide few examples of their behavior.Synthesis retains the host window’s residual background and substitutes one target-appliance power sequence.Source: 1) Training-Window Synthesis · p. 4
- Each mini-batch assigns fixed positions to sparse appliances, resampling segments and scaling modes until each assigned window meets the activation requirement.The quota fixes assigned positions, not the number of windows in which an appliance is active; other sparse appliances in the host remain as recorded.Source: 1) Training-Window Synthesis · p. 4
- An assigned activation segment may be rescaled in amplitude or duration before insertion, with duration scaling implemented by interpolation.Both targets follow from the substituted sequence, and partial-visibility semantics are used for insertion.Source: 1) Training-Window Synthesis · pp. 4–5
2) Label-Preserving Aggregate Recomposition:
- A recomposed pair uses an anchor produced by synthesis and replaces its residual background with one from another recorded source window.Both windows share the same synthesized target matrix.Source: 2) Label-Preserving Aggregate… · p. 5
- Every modeled appliance power sequence and its thresholded operating-state target are preserved pointwise, while the aggregates differ exactly by the residual-background change.This label preservation follows arithmetically from the construction.Source: 2) Label-Preserving Aggregate… · p. 5
- Replacement backgrounds are resampled if numerically invalid or if the recomposed aggregate exceeds the configured physical range; the pair is not claimed to represent the true joint distribution or a causal intervention.Sampling stops with an error if no valid replacement is found.Source: 2) Label-Preserving Aggregate… · p. 5
C. Prediction Consistency
Prediction consistency supplements full task supervision on label-preserving pairs with a penalty on power-prediction disagreement. Appliance-specific reliability thresholds and margins determine which disagreements are penalized and by how much.
C. Prediction Consistency
- Both windows receive complete power and operating-state supervision, while consistency penalizes disagreement between corresponding per-appliance power predictions only when both meet a fixed reliability criterion.The penalty is applied to gated power outputs, not to the representations producing them.Source: C. Prediction Consistency · p. 5
1) Reliability Gate and Margin:
- A reliability threshold γ_k determines whether a pair enters the term for appliance k, and a margin ε_k sets how much disagreement is tolerated.Both appliance-specific thresholds are fixed in advance.Source: 1) Reliability Gate and Margin · p. 5
- The consistency predictions come from a second evaluation of the paired batch with dropout disabled, while ordinary task losses use the standard training pass.Both prediction windows remain differentiable; only the binary gate is detached.Source: 1) Reliability Gate and Margin · p. 5
- A pair is gated in only when both window predictions lie within γ_k of their common target, and disagreement is penalized only beyond ε_k.Only the combination of a passed gate and excess disagreement carries a penalty.Source: 1) Reliability Gate and Margin · p. 6, Figure 3 · p. 6
- The thresholds are calibrated before consistency training from dropout-disabled predictions of checkpoints trained without the term, then held fixed throughout training.The reliability threshold uses a quantile of maximum pair error; the margin uses a quantile of disagreement among pairs passing the gate.Source: 1) Reliability Gate and Margin · p. 6
2) Output Consistency:
Output consistency averages gated excess disagreement across pairs and appliances, while the overall objective combines it with equally weighted task losses on both windows. The recomposition and consistency components are used only during training.
2) Output Consistency:
- The consistency term averages gated excess disagreement over pairs and then appliances, so an appliance with many gate-open pairs does not dominate one with few.Pairs closed by the gate remain in the denominator and contribute zero to the numerator.Source: 2) Output Consistency · p. 6
- Disagreement between paired predictions cancels any additive pointwise error component shared by both windows, but the full gated loss still depends on each window’s mean absolute error.The resulting gradient can affect the complete upstream computation, including the shared stage.Source: 2) Output Consistency · p. 6
D. Overall Training Objective
- Every component of the task loss applies to both windows with equal weighting, and the consistency gradient is accumulated before one clipping operation and optimizer step.The task-loss gradient is computed first, followed by the dropout-disabled consistency pass.Source: D. Overall Training Objective · p. 6
- At inference, the method uses the standard single-window path without recomposed windows, fixed thresholds, or the consistency term, adding no inference-time module or parameter.These components are used during training only.Source: D. Overall Training Objective · p. 7
IV. EXPERIMENTS AND RESULTS
The experiments evaluate multi-appliance NILM on three public datasets, comparing methods on households excluded from training. Performance is measured per appliance with power-error and operating-state metrics, plus unweighted macro averages.
IV. EXPERIMENTS AND RESULTS
- The evaluation compares complete-method FLAME with single-window FLAME and compares FLAME with four alternative multi-appliance architectures under a common protocol.Source: §IV · p. 7
A. Datasets and Evaluation Metrics
- REDD and UK-DALE use 6-s data and REFIT uses 8-s data, with training and test households disjoint in each dataset.The test household contributes no training data; signals are evaluated in 720-sample windows with a 120-sample stride.Source: A. Datasets and Evaluation Metrics · p. 7
- Each appliance is evaluated by MAE, SAE, and operating-state F1, while macro values are unweighted means over appliances.The operating-state probability threshold is 0.3.Source: A. Datasets and Evaluation Metrics · p. 7
B. Experimental Setup
The experimental setup compares the complete recomposition-and-consistency method with single-window FLAME, alongside a separate architecture comparison. On unseen households, the proposed method lowers macro MAE and SAE relative to single-window FLAME across all three datasets.
B. Experimental Setup
- The baseline trains FLAME on single windows with task loss alone; the proposed variant uses label-preserving pairs, complete task loss on both windows, and λcons = 0.8.Values are averaged over three independent trials; macro-MAE comparisons also report sample standard deviation and paired-run counts.Source: B. Experimental Setup · p. 7
- The separate architecture experiment evaluates five models under a shared protocol without pair construction, consistency loss, or architecture-specific hyperparameter tuning.BERT4NILM uses the common protocol rather than its original masked training, and MA-NILM is adapted to sequence-to-sequence prediction.Source: B. Experimental Setup · p. 7, Table II · p. 8
C. Experimental Results
- Macro MAE falls from 14.75 to 13.14 W on REDD, from 8.88 to 8.51 W on UK-DALE, and from 15.83 to 14.55 W on REFIT versus single-window FLAME.Table I reports results on the unseen test household for each dataset; macro SAE also falls on all three datasets.Source: C. Experimental Results · pp. 7–8, Table I · p. 8
D. Architecture Comparison
Under a common training protocol, FLAME has the lowest macro MAE among the compared architectures on all three datasets. The conclusion also reports lower macro MAE for the proposed training method, while bounding the evidence to three datasets and one unseen household per dataset.
D. Architecture Comparison
- FLAME attains the lowest macro MAE among the five compared architectures on all three datasets under the common protocol.Because alternatives were reimplemented without architecture-specific tuning, the ranking may reflect both architecture and compatibility with the shared setting.Source: D. Architecture Comparison · p. 8
V. CONCLUSION
- The method recomposes aggregates to change only residual background while preserving target power and operating-state labels pointwise, then constrains prediction disagreement under reliability and margin thresholds.It uses FLAME's two-stage shared-to-specific expert routing; deployment adds no inference-time module or parameter.Source: §V · p. 8
- On REDD, UK-DALE, and REFIT, the proposed method lowers macro MAE from 14.75 to 13.14 W, from 8.88 to 8.51 W, and from 15.83 to 14.55 W versus single-window training.Source: §V · p. 8
- Method-comparison evidence covers three datasets, one unseen household per dataset, and one fixed consistency weight.The recomposition's label preservation is a property of its construction, not evidence about the true joint distribution or a physical intervention.Source: §V · p. 8