PaperlayerView exact revision on arXiv

Source-linked AI summary

Transfer Learning for Non-Intrusive Load Monitoring

Michele DIncecco, Stefano Squartini, Mingjun Zhong

arXiv:1902.08835v3cs.LGstat.ML

TL;DR

NILM must recover appliance loads from aggregate mains despite non-identifiability, and seq2point learning had mainly been evaluated within the same data domain. This paper proposes appliance and cross-domain transfer learning, finding that appliance features transfer and that domain similarity determines whether fine tuning is needed.

  • Problem

    Seq2point learning had been evaluated only within the same domain, leaving its generalisation to different appliances and household-data domains unclear.

  • Method

    The paper evaluates appliance transfer learning and cross-domain transfer learning for seq2point models, reusing learned CNN features and selectively fine tuning fully connected layers.

  • Results

    Washing-machine CNN features transfer to other appliances; seq2point models can be applied without fine tuning in similar domains, whereas different domains benefit from fine tuning fully connected layers.

  • Takeaways & Limitations

    Transfer learning may support reusable residential-appliance models, reduced sensor requirements, and computational savings through reuse of pretrained features.

Abstract

from arXiv · show

Non-intrusive load monitoring (NILM) is a technique to recover source appliances from only the recorded mains in a household. NILM is unidentifiable and thus a challenge problem because the inferred power value of an appliance given only the mains could not be unique. To mitigate the unidentifiable problem, various methods incorporating domain knowledge into NILM have been proposed and shown effective experimentally. Recently, among these methods, deep neural networks are shown performing best. Arguably, the recently proposed sequence-to-point (seq2point) learning is promising for NILM. However, the results were only carried out on the same data domain. It is not clear if the method could be generalised or transferred to different domains, e.g., the test data were drawn from a different country comparing to the training data. We address this issue in the paper, and two transfer learning schemes are proposed, i.e., appliance transfer learning (ATL) and cross-domain transfer learning (CTL). For ATL, our results show that the latent features learnt by a `complex' appliance, e.g., washing machine, can be transferred to a `simple' appliance, e.g., kettle. For CTL, our conclusion is that the seq2point learning is transferable. Precisely, when the training and test data are in a similar domain, seq2point learning can be directly applied to the test data without fine tuning; when the training and test data are in different domains, seq2point learning needs fine tuning before applying to the test data. Interestingly, we show that only the fully connected layers need fine tuning for transfer learning. Source code can be found at https://github.com/MingjunZhong/transferNILM.

I. INTRODUCTION

NILM recovers appliance energy consumption from mains readings, avoiding costly per-appliance sensors, but the source-separation problem is non-identifiable. Deep learning has achieved strong NILM performance, motivating transfer-learning approaches and broader evaluation.

  • I. INTRODUCTION: NILM recovers each appliance’s energy consumption using only household mains readings instead of dedicated appliance sensors.Per-appliance sensing is expensive to install and difficult to maintain.
  • I. INTRODUCTION: NILM feedback can help householders understand energy use and has been reported to reduce household energy consumption by 15%.The same information can support smart-grid optimisation and consumption-based tariffs.
  • I. INTRODUCTION: Deep learning approaches have achieved state-of-the-art performance in NILM as household energy datasets expand across countries.The cross-country availability of data motivates studying whether learned models generalise beyond their training domain.
  • I. INTRODUCTION: The paper examines whether seq2point learning can transfer across appliances and data domains rather than only being evaluated within the same domain.It proposes appliance transfer learning and cross-domain transfer learning.
  • I. INTRODUCTION: NILM is a non-identifiable single-channel blind source-separation problem because multiple appliance sources must be recovered from one mains observation.Domain knowledge has therefore been incorporated into NILM models to alleviate identifiability issues.

1) Unsupervised learning:

Unsupervised NILM commonly models appliance states with factorial hidden Markov models, while supervised learning uses paired mains and appliance readings to learn a nonlinear mapping. Deep neural networks can learn appliance features directly without manually extracting them.

  • 1) Unsupervised learning:: AFHMM models NILM by treating each appliance as a hidden Markov model whose power readings add to the mains.Its objective is to infer the most probable appliance states, although the problem remains non-identifiable.
  • 1) Unsupervised learning:: AFHMM performance can be improved by imposing local and global appliance constraints derived from domain knowledge.Examples include power levels, ON-OFF transitions, duration, total energy, cycle counts, and occupancy time.
  • 1) Unsupervised learning:: Manual extraction of domain knowledge makes these constrained methods difficult to use.
  • 2) Supervised learning:: Supervised NILM learns a function f mapping mains readings Y to appliance readings X from paired observations.This formulation treats NILM as a nonlinear regression problem.
  • 2) Supervised learning:: Deep neural networks can learn appliance power levels, ON-OFF changes, and activity duration without manually extracting those features.CNNs were reported as performing best among the listed neural-network architectures.
  • 2) Supervised learning:: DNN-based NILM treats unknown appliances, base load, and model noise as background while extracting the target appliance.This makes the approach resemble a denoising technique.

III. METHODS

The paper investigates whether seq2point learning transfers across appliances and domains. Seq2point predicts an appliance’s midpoint power from a sliding mains window, and transfer experiments compare direct reuse with limited fine tuning.

  • III. METHODS: The paper studies appliance transfer learning and cross-domain transfer learning for seq2point models.Prior work had mainly assessed deep NILM methods within the same domain, with limited generalisability evidence.
  • A. Sequence-to-point learning: Seq2point maps a mains window Y_t:t+W−1 to the midpoint appliance value x_τ of the corresponding output window.The midpoint is expected to depend on mains information before and after it.
  • A. Sequence-to-point learning: The model represents the midpoint prediction as x_τ = f(Y_t:t+W−1)+ϵ.The network is trained to learn the nonlinear relationship between the mains window and the target midpoint.
  • A. Sequence-to-point learning: Zero padding of W/2 samples at both sequence endpoints allows seq2point inference across the full input sequence.The architecture produces one prediction for every appliance sample rather than averaging predictions from multiple windows.
  • A. Sequence-to-point learning: Earlier seq2point evaluations used UK-DALE and REDD with training and testing performed within the same domain, leaving generalisation unclear.The paper therefore evaluates appliance and cross-domain transfer.

B. Appliance transfer learning

The transfer experiments test whether learned appliance signatures can be shared and whether seq2point models can move between household-data domains. The approach reuses CNN features and tunes only fully connected layers when adaptation is used.

  • B. Appliance transfer learning: Appliance transfer learning asks whether signatures learned from one appliance can transfer to other appliances.The signatures include ON-OFF changes, power levels, and activity duration.
  • B. Appliance transfer learning: Shared feature channels are motivated by similar learned signatures across appliances, potentially reducing training costs and sensor requirements.
  • B. Appliance transfer learning: CNN layers trained on washing machines are reused for other appliances, while only the fully connected layers are trained for the new appliance.The washing-machine CNN is selected as the transferable feature extractor.
  • B. Appliance transfer learning: Cross-domain transfer learning trains a model in one region or country and tests it in another while seeking to use as little test-domain information as possible.
  • B. Appliance transfer learning: The experiments compare direct application of a pretrained model with fine tuning on a small test-domain subset, without tuning CNN layers.Only the fully connected layers are tuned because CNN tuning did not improve performance in the experiments.
  • A. Data sets: REFIT, UK-DALE, and REDD provide household electricity data from different countries and sampling frequencies for the transfer experiments.Sampling-frequency differences require preprocessing to align readings before NILM algorithms are applied.

1) REFIT:

The study uses REFIT as a large UK training domain and compares it with UK-DALE and REDD, while seq2point architectures visualize learned appliance features.

  • REFIT contains active-power readings from 20 English buildings collected between 2013 and 2015 at 8-second intervals.
  • REFIT is used to train deep learning models because it is the largest of the three datasets studied.The authors expected its large volume of readings to support generalization to unseen houses.
  • UK-DALE covers five UK buildings from 2013 to 2015, with mains sampled every 1 second and appliances every 6 seconds.
  • Figure 2 compares appliance inputs and ground truth with last-convolutional-layer outputs from seq2point models trained on five appliance types.
  • REDD contains measurements from six US buildings, including mains sampled every 1 second and appliances sampled every 3 seconds.Its observations lasted between 3 and 19 days, and high-frequency current and voltage data were also available.

4) Preparing training and test data:

Training and test data are cleaned, aligned, normalized, and organized across appliance and domain-transfer experiments using the REFIT sampling interval.

  • Large missing-data segments are removed, while shorter gaps are retained as potential noise for regularization.
  • UK-DALE buildings 1 and 2 provide about 4.01 million samples over roughly 12 months, and appliance data are down-sampled from 6 to 8 seconds.
  • REDD experiments use microwave, fridge, dish washer, and washing machine data from houses 1–3, with about 1.2 million samples per appliance.The dataset does not contain kettle recordings.
  • For appliance transfer, CNN layers from appliance A are transferred and frozen while only appliance B’s fully connected layers are trained.
  • Data are normalized using appliance- or mains-specific means and standard deviations before being fed into the models.The normalization statistics are used only for preprocessing, not to inform appliance means or variances.

B. Metrics

The evaluation uses four error metrics covering pointwise prediction, total energy, daily energy, and squared disaggregation error.

  • Mean absolute error (MAE) averages the absolute difference between predicted and ground-truth appliance power at each time point.
  • Normalized signal aggregate error (SAE) measures relative error in an appliance’s total energy consumption.It compares predicted total energy with ground-truth total energy.
  • Energy per day (EpD) measures absolute error in predicted energy usage over a day period.The metric is intended for settings where users care about total energy consumed over a period.
  • Normalized disaggregation error (NDE) measures normalized error from the squared difference between predicted and ground-truth appliance signals.

C. Settings for training neural networks

Seq2point models map fixed-length mains windows to appliance midpoints, use early stopping during training, and are evaluated both within and across domains.

  • Settings for training neural networks: Each seq2point input is a 599-point aggregate active-power window, with the corresponding appliance midpoint as the target.Other window lengths could also be evaluated experimentally.
  • Settings for training neural networks: Training uses TensorFlow, the ADAM optimizer, and early stopping to reduce overfitting.Dropout was not used in these experiments but was identified as a possible future development.
  • Motivation: The paper investigates whether seq2point learning generalizes across domains because earlier evaluations were primarily conducted within the same domain.
  • Experimental settings: The paper evaluates same-domain training and testing on REFIT, UK-DALE, and REDD, using AFHMM as a baseline.
  • Cross-domain transfer: Training on REFIT and testing on UK-DALE improves seq2point performance for MAE and SAE, whereas testing on REDD does not.The authors relate this contrast to the similar UK domain of REFIT and UK-DALE versus REDD’s different domain.

1) Appliance transfer learning:

Appliance transfer learning reuses CNN features trained on washing-machine data while retraining only fully connected layers for other appliances. Its results are comparable to standard appliance-specific training.

  • Appliance transfer learning: Washing-machine CNN layers were reused for kettle, microwave, fridge, and dishwasher transfer learning.Only the fully connected layers were trained for each target appliance.
  • Appliance transfer learning: ATL was evaluated with CNN layers trained on REFIT using washing-machine data and tested on REFIT.
  • Appliance transfer learning: ATL predictions were comparable to standard training across the evaluated appliances.The comparison used appliance-specific seq2point models and models with transferred CNNs.
  • Appliance transfer learning: Using washing-machine CNN layers for all appliances could reduce the need to collect appliance-specific data and hardware.The authors associate this strategy with reduced hardware cost.

2) Cross-domain transfer learning:

Cross-domain transfer learning tests seq2point models trained on one dataset and applied to other datasets, with fine tuning used to adapt the model when needed. Similar domains support direct transfer, whereas different domains benefit from fine tuning.

  • Cross-domain transfer learning: CTL trains neural networks on a large dataset and uses them as pretrained models for other datasets, where they are fine tuned in the new domain.
  • Cross-domain transfer learning: REFIT-trained seq2point models performed similarly on REFIT and UK-DALE, suggesting direct application to UK-DALE.
  • Cross-domain transfer learning: REFIT-trained seq2point performance greatly decreased on REDD compared with training on REDD, indicating that direct transfer was unsuccessful there.The authors relate this to different appliance signatures across countries.
  • Cross-domain transfer learning: Fine tuning improved seq2point performance on REDD but decreased it on UK-DALE.The authors note that UK-DALE and REFIT have similar domains, while REDD differs from REFIT.
  • Cross-domain transfer learning: All evaluated transfer learning strategies provided reasonable appliance-energy predictions across the datasets.
  • Cross-domain transfer learning: Different domains require fine tuning for improved seq2point performance, whereas similar domains may not require it.

VI. CONCLUSIONS

The paper evaluates transferability of seq2point learning across appliances and datasets using ATL and CTL. It concludes that washing-machine features transfer to other appliances, while fine tuning depends on domain similarity.

  • Conclusions: The paper investigates transfer learning for NILM because prior seq2point studies evaluated training and testing only within the same domain.
  • Conclusions: The authors hypothesize that CNN-extracted appliance features are invariant across appliances and data domains, and report experiments supporting this hypothesis.
  • Conclusions: Washing-machine CNN layers can be applied to other appliances, after which fully connected layers are trained on unseen data.
  • Conclusions: Similar training and testing domains may not require fine tuning, whereas different domains benefit from fine tuning to adapt the model.
  • Conclusions: Transfer learning may support a shared residential-appliance model, reduce appliance-specific sensors, and save computation through reuse of pretrained features.
Loading 1902.08835v3…