Elsevier

Energy and Buildings

Volume 158, 1 January 2018, Pages 1461-1474
Energy and Buildings

Denoising autoencoders for Non-Intrusive Load Monitoring: Improvements and comparative evaluation

https://doi.org/10.1016/j.enbuild.2017.11.054Get rights and content

Highlights

  • NILM is performed by using a denoising autoencoder architecture.

  • Several improvements have been introduced to the traditional disaggregation scheme with Deep Neural Networks.

  • The performance has been evaluated on three publicly available datasets.

  • Tests have been conducted on noised and denoised scenario, in seen and unseen conditions.

  • On average, the proposed approach outperforms the AFAMAP algorithm.

Abstract

Non-Intrusive Load Monitoring (NILM) is the task of determining the appliances individual contributions to the aggregate power consumption by using a set of electrical parameters measured at a single metering point. NILM allows to provide detailed consumption information to the users, that induces them to modify their habits towards a wiser use of the electrical energy. This paper proposes a NILM algorithm based on the Deep Neural Networks. In particular, the NILM task is treated as a noise reduction problem addressed by using denoising autoencoder (dAE) architecture, i.e., a neural network trained to reconstruct a signal from its noisy version. This architecture has been initially proposed by Kelly and Knottenbelt (2015), and here is extended and improved by conducting a detailed study on the topology of the network, and by intelligently recombining the disaggregated output with a median filter. An additional contribution of this paper is an exhaustive comparative evaluation conducted with respect to one of the reference work in the field of Hidden Markov Models (HMM) for NILM, i.e., the Additive Factorial Approximate Maximum a Posteriori (AFAMAP) algorithm. The experiments have been conducted on the AMPds, UK-DALE, and REDD datasets in seen and unseen scenarios both in presence and in absence of noise. In order to be able to evaluate AFAMAP in presence of noise, an HMM model representing the noise contribution has been introduced. The results showed that the dAE approach outperforms the AFAMAP algorithm both in seen and unseen condition, and that it exhibits a significant robustness in presence of noise.

Introduction

Non-Intrusive Load Monitoring (NILM) is the task of extracting the energy consumed by individual appliances from a single metering point [1], [2], [3]. Indeed, several studies demonstrated that providing detailed appliance consumption information can lead to savings greater than 12% [1], [4], [5], [6], [7], and NILM provides this information without requiring dedicated sensors for each appliance. This allows the reduction of installation costs and the level of intrusiveness for the measurement, thus augmenting the chance of a large scale acceptance by the users. The estimated energy savings are a consequence of several factors that involve both the residential users and the energy providers. Regarding the users, detailed appliance consumption information would allow them to take the proper actions for reducing their bills, e.g., by replacing the inefficient appliances with more efficient ones. Energy providers, on the other hand, can exploit this information in order to predict the energy demand, to apply advanced management policies, and to prevent overloading or blackouts over the energy network [8].

The research on NILM has been particularly active in the last years, and many different approaches have been proposed. Nonetheless, a general framework for NILM can be defined and it comprises a feature extraction stage followed by a load identification stage [2]. In the literature, a first classification criterion regards the feature extraction stage and it divides the algorithms based on steady state features from the ones based on transient state features [2], [9], [10], [11]. In the former, features are extracted from the signals after an appliance has completed a state transition. In the latter, features are extracted during a state transition. A second criterion regards the requirement of user intervention for creating appliance models, and it divides supervised from unsupervised approaches [12], [13]. The latter have been the preferred choice in the literature, since they represent the most convenient approach for end-users. On the algorithmic side, machine learning techniques have been largely employed since they exhibited noteworthy disaggregation performance: solutions have been proposed that are based on Factorial Hidden Markov Models (FHMMs) [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], Neural Networks (NN) [26], [27], [28], [29], [30], graph-based signal processing [31], Support Vector Machines (SVM) [32], k-Nearest Neighbours [32], and Decision Trees [33]. For a recent review and a taxonomy, please refer to [2], [10], [34].

Among the techniques appeared in the literature, Deep Neural Networks (DNN) have been devoting particular attention in the last years, since they exhibited noteworthy performance for load disaggregation [27], [28], [29]. In [28], the authors proposed an approach based on Long Short-Term Memory (LSTM) neural networks [35]. The algorithm consists in training a neural network for each appliance in order to predict a sample of the disaggregated active power from a segment of aggregated data. Neural networks have been combined with HMMs in [29]: the emission probabilities of the HMM are modelled by a Gaussian distribution for state representing the single load, and by a DNN for state representing the aggregated signal. Similarly to [28], LSTMs have been also employed in [27], this time combined with convolutional layers at the input of the network to extract the features of the signal directly from raw data. In the same paper, NILM is treated also as a noise reduction problem, where the clean signal is represented by the disaggregated appliance profile, and the noise signal by the remaining profiles and the measurement noise. Noise reduction is performed by using a denoising autoencoder (dAE) composed of convolutional and fully connected layers that estimate the appliance profile from the aggregated noisy signal. An additional approach proposed in [27] uses a neural network that estimates the start time, the end time, and the mean power demand of each appliance. In the experiments conducted by the authors on the UK recording Domestic Appliance-Level Electricity dataset (UK-DALE) [36], they demonstrated that the most performing approach is represented by the dAE network, that outperformed both the other DNN architectures, and the FHMM method proposed in [25].

In this paper, several algorithmic and architecture improvements to the dAE approach for NILM are proposed and an exhaustive comparative evaluation with the AFAMAP (Additive Factorial Approximate Maximum a Posteriori) algorithm [17] is conducted. In particular, compared to [27] the dAE approach for load disaggregation is improved by conducting a detailed study on the topology of the network, and by introducing pooling and upsampling hidden layers, and the rectifier linear unit (ReLU) activation function [37] in the output layer. Additionally, the network output is recombined by using a median filter on the overlapped portions of the disaggregated signal. The second contribution is an exhaustive performance comparison between AFAMAP and the dAE approach. Indeed, FHMMs have been largely employed in the last years since they are an effective approach for load disaggregation, and AFAMAP, in particular, received noteworthy attention by the scientific community [38], [39], [40]. However, up to the authors’ knowledge, an exhaustive performance comparison between the two methods has not been yet conducted, and it is authors’ opinion that it can provide an useful reference for the scientific community. Indeed, the authors of [27] compare their proposed approaches to the FHMM method implemented in NILMTK [41], but their comparison does not consider more advanced FHMM algorithms such as AFAMAP [17]. Additionally, their experiments consider only a noised scenario on a single dataset (UK-DALE). Here, the evaluation is performed on three datasets, UK-DALE [36], AMPds [42], and REDD [25] in different conditions: firstly, the algorithms are evaluated on denoised and noised scenarios. In the denoised scenario, the aggregated signal is the sum of the power profiles of the appliances that are disaggregated. In the noised scenario, the aggregated signal comprises also measurement noise and the contributions of unknown appliances. Successively, the algorithms generalisation capabilities are evaluated by performing disaggregation on the data acquired in a house not considered in the training phase (unseen scenario). The performance is evaluated by using both energy-based metrics and state-based metrics [41]: the first, evaluate the capability of the algorithm to estimate the actual power profile of the appliances, while the second the capability of estimating whether the appliance is in the “on” or “off” state. In order to perform the experiments in presence of noise, a Rest of the World (RoW) model has been introduced in the original AFAMAP [17] algorithm. This model represents all the appliances but the ones of interest and makes AFAMAP able to operate in a noised scenario. The obtained results show that on average the dAE approach outperforms AFAMAP in all the addressed experimental conditions.

The outline of the paper is the following: Section 2 introduces the general problem of NILM. Section 3 presents the Deep Neural Networks approach and the advancements introduced with respect to [27]. Section 4 describes the AFAMAP algorithm and the approach adopted for disaggregation in the noisy scenario. In Section 5, the experimental procedure and the obtained results are presented and discussed. Finally, Section 6 concludes the paper and proposes future developments.

Section snippets

Problem statement

The NILM problem can be formulated as follows: let y(t) be the aggregated signal measured at the time index t. Without lack of generality, here we suppose that y(t) represents the active power. y(t) can be expressed as the sum of the active power contributions of each appliance:y(t)=i=1Nyi(t)+e(t),where N is the number of appliances, yi(t) is the individual contribution of appliance i, and e(t) is a noise term. The NILM problem is, thus, the task of finding the individual appliance

Deep Neural Networks based algorithm

The NILM task can formulated as a denoising problem by expressing the aggregated signal as the sum of the power consumption of the appliance of interest and a noise component that incorporates all the remaining contributions. In particular, Eq. (1) can be reformulated as:y(t)=yj(t)+vj(t),for j = 1, 2, …, N, wherevj(t)=i=1ijNyi(t)+e(t),represents an overall noise term for the appliance j that comprises both the measurement noise and the contributions of the other appliances. Thus, for obtaining y

Factorial Hidden Markov Models based approach

FHMMs have been introduced in [47] as an extension of HMMs to model time series that depend on multiple hidden processes. Starting from the work of Kim and colleagues [14], FHMMs have been largely employed for NILM and several approaches have been proposed in the literature [15], [18], [20], [23], [39], [48], [49]. Among them, AFAMAP [17] represents an effective algorithm able to achieve high performance with a reasonable computational cost.

In this section, the HMM appliance model and the

Experiments

This section describes the experiments conducted to evaluate the performance of the dAE approach and of the AFAMAP algorithm. Firstly, the performance metrics, the datasets, and the experimental procedure are described. Then, the obtained results are presented and discussed.

Conclusion

In this paper, a DNN architecture based on the denoising autoencoder topology has been proposed. Compared to the work by Kelly and Knottenbelt [27] several improvements have been introduced. In the training phase, the variable step size has been adopted, with an early stopping criterion based on the performance metric calculated on the validation set. In the disaggregation phase, the median filter has been applied to combine the overlapped portion of signal in the sliding window analysis of the

Acknowledgements

We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources and support, and Netribe Business Solution SRL for supporting this research.

References (54)

  • A. Zoha et al.

    Non-intrusive load monitoring approaches for disaggregated energy sensing: a survey

    Sensors

    (2012)
  • Z. Wang et al.

    Residential appliances identification and monitoring by a nonintrusive method

    IEEE Trans. Smart Grid

    (2012)
  • K. Ehrhardt-Martinez et al.

    Advanced Metering Initiatives and Residential Feedback Programs: A Meta-Review for Household Electricity-Saving Opportunities

    (2010)
  • J. Laitner et al.

    Examining the scale of the behaviour energy efficiency continuum

  • G. Gardner et al.

    The short-list: the most effective actions us households can take to curb climate change

    Environ. Sci. Policy Sustain. Environ.

    (2008)
  • L.D. Baets et al.

    On the Bayesian optimization and robustness of event detection methods in NILM

    Energy Build.

    (2017)
  • M. Zeifman et al.

    Nonintrusive appliance load monitoring: review and outlook

    IEEE Trans. Consum. Electron.

    (2011)
  • R. Bonfigli et al.

    Unsupervised algorithms for non-intrusive load monitoring: an up-to-date overview

    2015 IEEE 15th International Conference on Environment and Electrical Engineering (EEEIC)

    (2015)
  • H. Kim et al.

    Unsupervised disaggregation of low frequency power measurements

  • O. Parson et al.

    A scalable non-intrusive load monitoring system for fridge-freezer energy efficiency estimation

    Proc. of the 2nd Int. Workshop on Non-Intrusive Load Monitoring

    (2014)
  • J. Kolter et al.

    Approximate inference in additive factorial HMMs with application to energy disaggregation

    J. Mach. Learn. Res.

    (2012)
  • I. Valera et al.

    Infinite factorial unbounded-state hidden markov model

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2016)
  • Y. Li et al.

    Energy disaggregation via hierarchical factorial hmm

    Proc. of the 2nd Int. Workshop on Non-Intrusive Load Monitoring

    (2014)
  • M. Zhong et al.

    Signal aggregate constraints in additive factorial HMMs with application to energy disaggregation

    Adv. Neural Inf. Process. Syst.

    (2014)
  • M. Zhong et al.

    Interleaved factorial non-homogeneous hidden markov models for energy disaggregation

  • S. Makonin et al.

    Exploiting HMM sparsity to perform online real-time nonintrusive load monitoring

    IEEE Trans. Smart Grid

    (2016)
  • M. Johnson et al.

    Bayesian nonparametric hidden semi-Markov models

    J. Mach. Learn. Res.

    (2013)
  • Cited by (151)

    View all citing articles on Scopus
    View full text