PaperlayerView exact revision on arXiv

Source-linked AI summary

Subtask Gated Networks for Non-Intrusive Load Monitoring

Changho Shin, Sunghwan Joo, Jaeryun Yim, Hyoseop Lee, Taesup Moon, Wonjong Rhee

arXiv:1811.06692v1cs.LGstat.APstat.ML

TL;DR

NILM estimates individual-appliance electricity use from a household’s aggregate consumption, supporting applications such as consumer feedback and utility demand management. This paper combines regression with on/off classification in a subtask gated network, whose performance is reported as 15–30% better than the state-of-the-art deep-learning solution on REDD and UK-DALE. The authors identify regression problems with inherent on/off states as a promising scope for the approach.

  • Problem

    NILM estimates individual-appliance power from aggregate electricity consumption, with energy-disaggregation information used for consumer feedback and utility demand management.

    Source: Introduction · p. 1

  • Method

    The subtask gated network multiplies a regression network’s output by an on/off-classification network’s probability and uses classification loss to train the gate.

    Source: Introduction · p. 1

  • Results

    The proposed solution performed 15–30% better than the state-of-the-art deep-learning solution on REDD and UK-DALE using MAE and SAEδ.

    Source: Conclusions · p. 7

  • Takeaways & Limitations

    The results indicate that combining regression and classification subnetworks is a promising direction for regression tasks with inherent on/off states.

    Source: Conclusions · p. 7

    The method handles only appliance on/off states, while some appliances have multiple states.

    Source: Conclusions · p. 7

Abstract

from arXiv · show

Non-intrusive load monitoring (NILM), also known as energy disaggregation, is a blind source separation problem where a household's aggregate electricity consumption is broken down into electricity usages of individual appliances. In this way, the cost and trouble of installing many measurement devices over numerous household appliances can be avoided, and only one device needs to be installed. The problem has been well-known since Hart's seminal paper in 1992, and recently significant performance improvements have been achieved by adopting deep networks. In this work, we focus on the idea that appliances have on/off states, and develop a deep network for further performance improvements. Specifically, we propose a subtask gated network that combines the main regression network with an on/off classification subtask network. Unlike typical multitask learning algorithms where multiple tasks simply share the network parameters to take advantage of the relevance among tasks, the subtask gated network multiply the main network's regression output with the subtask's classification probability. When standby-power is additionally learned, the proposed solution surpasses the state-of-the-art performance for most of the benchmark cases. The subtask gated network can be very effective for any problem that inherently has on/off states.

Introduction

NILM estimates individual-appliance power from aggregate electricity consumption, enabling applications such as consumer feedback and utility demand management. This work uses appliance on/off states in a subtask gated network and reports improved performance on REDD and UK-DALE.

Introduction

  • NILM estimates individual-appliance power using aggregate electricity consumption as its only input, avoiding separate measurement devices on each appliance.The paper also calls this process energy disaggregation.Source: Introduction · p. 1
  • NILM-derived energy information can provide consumer feedback and support utility actions, including reducing peak-hour usage in homes running particular appliances.The paper cites a reported 15% energy saving from consumer feedback (Neenan, Robinson, and Boisvert 2009).Source: Introduction · p. 1
  • The subtask gated network combines a regression DNN with an on/off-classification DNN, multiplying the regression output by the classification probability to form the final estimate.Its classification loss explicitly reflects classification errors and makes the classification subnetwork an on/off gate.Source: Introduction · p. 1
  • SGN improves performance by 15–30% on average in REDD and UK-DALE, while standby-power and hard-gating variants add improvements for some appliances.The introduction reports the 15–30% result before describing the variants’ additional gains for some appliances.Source: Introduction · pp. 1–2

Related Works

Prior NILM work includes FHMM variants and deep-learning approaches, but the latter do not explicitly use appliance on/off states with end-to-end training. SGN differs from typical multitask learning by using auxiliary classification output as a multiplier, with a loss that directly trains the gate.

Non-intrusive load monitoring

  • FHMM variants remain common, while deep NILM approaches still lack end-to-end use of appliance on/off states.Prior deep-learning approaches also include CNNs, RNNs, denoising autoencoders, and a CNN/hidden semi-Markov model (Kelly and Knottenbelt 2015; Huss 2015).Source: Non-intrusive load monitoring · p. 2

Multi-task learning

  • Typical MTL shares lower-layer parameters across tasks (Zhang et al. 2014; Yu and Jiang 2016; Liu et al. 2015; Cheng, Fang, and Ostendorf 2015); SGN instead multiplies the main output by auxiliary classification output.The auxiliary output is used directly to calculate the main task’s final output, beyond indirect influence through shared parameters and losses.Source: Multi-task learning · p. 2

Gating mechanism

  • Unlike gates trained through final-output losses, SGN adds an auxiliary-label gating loss so the model learns directly from the on/off gating task.The paper contrasts this with examples where the gating network is not explicitly trained on the gating task itself.Source: Gating mechanism · p. 2

Preliminary

NILM estimates appliance-level power sequences from aggregate consumption; this section formulates that task and describes the partial-window CNN models used for estimation.

Problem formulation of energy disaggregation

  • The disaggregation task is to estimate the power-consumption sequences of selected appliances from aggregate consumption, which also includes unknown-appliance use and noise.The formulation represents aggregate consumption as the sum of appliance loads plus Gaussian noise, then separates selected appliances from unknown consumption.Source: Problem formulation of energy… · p. 2

Deep neural networks for energy disaggregation

  • Deep networks estimate appliance power from partial aggregate-consumption sequences, with additional input windows on both sides to retain context.The output covers the corresponding partial appliance sequence; the sequence length and added-window length are fixed in the paper.Source: Deep neural networks for energy… · pp. 2–3
  • The study uses Zhang’s CNN architecture and excludes RNNs after tried RNN models performed worse, even with longer training.The authors state that the network architecture itself is not the main focus of their work.Source: Deep neural networks for energy… · p. 3

Subtask Gated Networks (SGN)

The SGN uses an on/off classification subtask to gate appliance-power regression, trains the subnetworks jointly, and considers standby-power and hard-gating variants.

Auxiliary classification subtask for NILM

  • The on/off classifier’s probability gates the regression output through element-wise multiplication, making appliance state an explicit part of NILM estimation.Each classification output element represents the probability that the appliance is on.Source: Auxiliary classification subtask for… · p. 3

Loss function

  • The joint loss combines overall network and classification-subnetwork losses, while regression learns only when the classifier predicts an on state.If the classifier saturates at zero, regression gradients are also zero; minimizing the classification loss can later revive the classifier when regression outputs are nonzero.Source: Loss function · pp. 3–4

Variants of subtask gated networks for NILM

  • SGN-sp modifies the zero-off-power assumption to account for fixed standby power, represented by a learnable scalar.The paper notes that standby demand may be small but is not negligible because appliances spend a significant portion of time off.Source: SGN with standby power (SGN-sp) · p. 4
  • Hard SGN replaces probability gating with a 0-or-1 gate using a 0.5 threshold; its classification loss still uses the probability output.The paper also describes a corresponding Hard SGN-sp formulation.Source: Hard SGN · p. 4

Experiments

Experiments on REDD and UK-DALE compare SGN with prior methods and examine its gating behavior and sensitivity to the evaluation period. SGN variants improve benchmark performance, while the classification loss and choice of δ shape the observed results.

Experiment settings

  • Across REDD and UK-DALE, REDD training used houses 2–6 and tested house 1, while UK-DALE trained on houses 1 and 3–5 and tested house 2.The experiments selected four appliance types for REDD and five for UK-DALE; only the last week of published, preprocessed UK-DALE data was used.Source: Experiment settings · p. 4
  • MAE measures power-estimation error, while SAEδ compares ground-truth and estimated energy sums over time periods and is normalized by time rather than ground-truth consumption.The chosen hourly normalization avoids outliers from near-zero denominators and division by zero when an appliance is off.Source: Experiment settings · p. 5

Experiment results

  • The reported average favored SGN over hard SGNs, although hard SGNs still outperformed previous work; improvements were observed for fridges and microwaves.The authors note appliance-dependent selection among SGN and its variants.Source: Experiment results · pp. 5–6
  • SGN and SGN-sp achieved 15–30% lower average error than state-of-the-art; SGN-sp improved MAE in 9/9 cases and SGN in 7/9, while both improved SAE in 6/9.The comparison covers the REDD and UK-DALE benchmark tables.Source: Table 2 · p. 5, Experiment results · p. 5, Table 3 · p. 6
  • Including Lon concentrates on-probabilities near zero and one; without it, the classifier can act as a scaling factor instead of learning the intended on/off task.The fridge, dishwasher, and kettle probability distributions differ in ways consistent with their reported operating patterns.Source: Experiment results · pp. 6–7
  • For dishwashers, gating filters small noisy off-state estimates; for fridges and kettles, classification gates regression outputs affected by off-state or other-appliance noise.The kettle’s regression output can show large values when other appliances are turned on.Source: Figure 4 · p. 6, Experiment results · p. 6
  • SGN performed best across the examined δ range, and performance improved as δ increased, consistent with positive and negative errors having more opportunity to cancel.The evaluation considers different time horizons because applications may require hourly, daily, or monthly energy use.Source: Figure 7 · p. 7, Experiment results · p. 7

Conclusions

The proposed SGN improves NILM performance over a state-of-the-art deep-learning solution, while the method currently handles appliance on/off states and is proposed for regression problems with related discrete states.

Conclusions

  • 15–30% better performance than the state-of-the-art deep-learning solution was reported on REDD and UK-DALE using MAE and SAEδ.Source: Conclusions · p. 7
  • The method handles appliance on/off states, while appliances with multiple states may require more precise labels.Source: Conclusions · p. 7
  • The authors propose extending SGN to regression problems where an auxiliary class can describe discrete states related to the output.Source: Conclusions · p. 8

Appendix A. The details of architecture hyperparameters

The appendix lists architecture details for two SGN subnetworks and gives selected layer specifications, including a sigmoid activation in the classification subnetwork.

Appendix A. The details of architecture hyperparameters

  • SGN uses two subnetworks, and only the classification subnetwork has a sigmoid activation at its last fully connected layer.Source: Appendix A · p. 10
  • The listed convolutional layers include filter sizes 10, 8, and 6 with 30, 30, and 40 filters, followed by three size-5 layers with 50 filters each.Source: Appendix A · p. 10
  • One listed fully connected layer has 1024 neurons.Source: Appendix A · p. 10
  • The architecture listings also include convolutional layers with filter size 4, eight filters, and stride 1.Source: Appendix A · p. 10

Appendix B. Sensitivity to input and output lengths

Across the investigated input and output size pairs, SGN outperformed DAE and Seq2Seq. The sensitivity experiments also show that output length and added input-window length affect performance.

Appendix B. Sensitivity to input and output lengths

Appendix C. Training with Loutput only

The appendix compares training with only Loutput against training with Loutput + Lon and reports generally poorer performance for the single-loss setup.

Appendix C. Training with Loutput only

  • Training with only Loutput generally performed worse than training with Loutput + Lon in the same setup.Source: Appendix C · p. 10
  • The appendix identifies results tables for the Loutput-only setup on REDD and UK-DALE.Source: Table 1 · p. 11, Table 2 · p. 11
Loading 1811.06692v1…