Elsevier

Energy and Buildings

Volume 224, 1 October 2020, 110299
Energy and Buildings

Generating realistic building electrical load profiles through the Generative Adversarial Network (GAN)

https://doi.org/10.1016/j.enbuild.2020.110299Get rights and content

Highlights

  • Generative Adversarial Network is used to generate building electric load.

  • The generated load could capture both the general trend and stochastic dynamics.

  • The proposed approach was tested using real smart meter data.

  • Potential applications include electric load generation, anonymization, and anomaly detection.

Abstract

Building electrical load profiles can improve understanding of building energy efficiency, demand flexibility, and building-grid interactions. Current approaches to generating load profiles are time-consuming and not capable of reflecting the dynamic and stochastic behaviors of real buildings; some approaches also trigger data privacy concerns. In this study, we proposed a novel approach for generating realistic electrical load profiles of buildings through the Generative Adversarial Network (GAN), a machine learning technique that is capable of revealing an unknown probability distribution purely from data. The proposed approach has three main steps: (1) normalizing the daily 24-hour load profiles, (2) clustering the daily load profiles with the k-means algorithm, and (3) using GAN to generate daily load profiles for each cluster. The approach was tested with an open-source database – the Building Data Genome Project. We validated the proposed method by comparing the mean, standard deviation, and distribution of key parameters of the generated load profiles with those of the real ones. The KL divergence of the generated and real load profiles are within 0.3 for majority of parameters and clusters. Additionally, results showed the load profiles generated by GAN can capture not only the general trend but also the random variations of the actual electrical loads in buildings. The proposed GAN approach can be used to generate building electrical load profiles, verify other load profile generation models, detect changes to load profiles, and more importantly, anonymize smart meter data for sharing, to support research and applications of grid-interactive efficient buildings.

Introduction

The term “building electrical load profile” refers to the variation of the building's electrical load versus time.1 Building electrical load profile is influenced by both physical and behavioral factors [1]. Building load profile analytics have wide applications, such as identification of unnecessary waste [2], load forecasting [3], customer segmentation and demand-side management [4], demand response planning and pricing [5], long-term resource planning, and renewable energy integration [6]. Therefore, collecting and analyzing load profiles plays a significant role in enhancing our understanding of how best to integrate buildings with a changing electric grid and to finally reduce building carbon footprints.

Building load profiles can be collected by smart meter data, which are the data (mostly electrical power consumption) collected by the smart meter at fine temporal granularities [7]. A smart meter is an electronic device that records consumption of electric energy and communicates the information to a remote server. Compared with conventional electricity meters, the smart meter has two key characteristics: fine temporal granularity and real-time communication. The temporal granularity of smart meter data is usually 5 to 15 mins.

With the pervasive deployment of smart meters, fine-grained electricity consumption data are becoming increasingly available [8]. While the use of smart meter data is unlocking large benefits, it also raises data privacy and security concerns [9]. Partly due to those privacy concerns, utility companies are unwilling or unable to share customer meter data for general research and analysis, which is a major barrier to advancing smart meter data analytics. To address the privacy concerns, existing studies have proposed privacy-preserving approaches, such as data aggregation [10], clustering [11], principal component analysis [12], and wavelet-based representation [13]. However, those privacy-preserving methods suffer from reduced data resolution and unavoidable information loss.

As building electrical load profiles can yield useful information and have wide applications, many studies have been done to analyze building load profiles, and these dated back to as early as the 1980s [14]. Studies on building load profiles can be classified into descriptive and generative studies. Descriptive studies aim to describe the characteristics of building load profiles. Price [2] defined five statistics to characterize daily load profile: (1) near-peak load, (2) near-base load, (3) high-load duration, (4) rise time, and (5) fall time, and demonstrated how load profile analysis can be used to estimate demand response effectiveness.

Generative studies aim to generate load profiles that are as realistic as possible, and they can avoid the privacy concerns of directly using smart meter data. Additionally, the relatively high cost of smart meters prevents them from being deployed in some regions; this problem can be solved by using generative studies instead. The generated load profiles can be used to test or verify demand control techniques [15] and to evaluate demand response policies.

Two common approaches are used to generate realistic building load profiles. The first is through energy simulation using white-box models. The white-box approach is physics based, which simulates the energy consumption with detailed building energy models. To facilitate physics-based building energy simulation, detailed assumptions about the building’s physics (envelope, system efficiency), predicted occupant behavior [16], and electrical appliance schedules [17] need to be provided, which demands expertise and is labor intensive. To avoid the tedious efforts needed to collect data on building physical and behavioral factors for white-box simulation, some studies use schedules and assumptions from reference building models to generate load profiles [18]. For instance, using U.S. Department of Energy (DOE) reference models, the National Renewable Energy Laboratory (NREL) generated and open sourced hourly load profile data for 16 commercial building types on OpenEI.2 However, the assumptions proposed by the modelers or in the reference building model may not necessarily reflect the reality of actual buildings [19], resulting in a gap in the load profiles between the white-box simulation and the real building [20].

The second approach is to use black-box models to regress building energy consumption, with features such as demographics [21], energy price [22], local climate [23], and presence of end-use appliances (aka “conditional demand analysis”) [24]. The black-box model can be developed through a probabilistic method [15], regression [25], or a neural network [26]. However, the black-box approach usually does not address individual end uses [27], and more importantly, it is unable to provide high temporal-granularity load variations. Fig. 1 summarizes major approaches taken to collect or generate building load profiles and their constraints.

The major contribution of this study is to propose a data-driven approach to generate realistic building electrical load profiles using the Generative Adversarial Network (GAN). This new approach can generate load profiles at the individual building or household level.

As a data-driven black-box approach, this approach can save tedious work of making assumptions on building physics and occupant schedules, which are required by physics-based models. In terms of privacy concern, GAN retains important statistical information such as the dynamic and stochastic behaviors of building loads, while anonymizing user-sensitive information to protect users’ privacy.

The proposed approach can be used to generate or forecast building load profiles, as well as to anonymize collected smart meter data. By protecting customer privacy, it is expected that more building owners will be encouraged to share their data for research and analysis.

The remaining of this paper is organized as follows. Section 2 proposes a two-step approach to applying GAN to generate realistic building load profiles. Section 3 presents the result of testing our proposed approach on smart meter data from real buildings. Hyper-parameter tuning, potential applications, and limitations of this approach are discussed in Section 4. Section 5 offers conclusions.

Section snippets

Method

Fig. 2 presents the technical roadmap to generate building load profiles through GAN, which has three major steps: data preprocessing, load profile clustering, and GAN, which is discussed in detail in this section.

Load profile clusters

The first step of clustering is to select the optimal number of clusters. We used DBI to select the optimal number of clusters. Clustering configuration with a lower DBI is always preferred, as a low DBI indicates the distance between data points within the same cluster is small compared with the distance between different clusters. We ran the clustering algorithm with different cluster numbers and selected the optimal number of clusters that had a low DBI value. As the k-means algorithm is

Hyper-parameter tuning

Hyper-parameter tuning is especially challenging in training a GAN model because GAN is a dynamic system consisting of two neural networks, with adversarial loss functions. The typical approach to train a deep neural network gradient descent method normally consists of going downhill in a static loss landscape. But with GAN, the loss landscape is no longer static. Every step taken down the hill changes the entire landscape a little, which makes it more difficult to find the minimal point

Conclusions

Building electrical load profiles have wide applications in research and analysis of building operations, energy efficiency, and building-grid interactions. Current approaches to obtain these load profiles may be either time-consuming, unable to reflect the dynamic and stochastic behaviors of real buildings, or have privacy concerns. In this study, we proposed a novel approach to generate realistic electrical load profiles through the Generative Adversarial Network (GAN), an unsupervised

CRediT authorship contribution statement

Zhe Wang: Conceptualization, Methodology, Investigation, Data curation, Writing - original draft, Writing - review & editing. Tianzhen Hong: Conceptualization, Methodology, Writing - review & editing, Resources, Visualization, Project administration, Funding acquisition.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (42)

  • X. Luo et al.

    Electric load shape benchmarking for small- and medium-sized commercial buildings

    Appl. Energy

    (2017)
  • G. Chicco

    Overview and performance assessment of the clustering methods for electrical load pattern grouping

    Energy

    (2012)
  • I. Benítez et al.

    Dynamic clustering segmentation applied to load profiles of energy consumption from Spanish customers

    Int. J. Electr. Power Energy Syst.

    (2014)
  • P.J. Rousseeuw

    Silhouettes: A graphical aid to the interpretation and validation of cluster analysis

    J. Comput. Appl. Math.

    (1987)
  • P. Price, “Methods for Analyzing Electric Load Shape and its Variability,” Lawrence Berkeley National Laboratory,...
  • X. Liu et al.

    CITIESData: a smart city data management framework

    Knowl. Inf. Syst.

    (2017)
  • Y. Wang et al.

    Sparse and redundant representation-based smart meter data compression and pattern extraction

    IEEE Trans. Power Syst.

    (2017)
  • S. Joseph et al.

    Real-time retail price determination in smart grid from real-time load profiles

    Int. Trans. Electr. Energy Syst.

    (2018)
  • N. A. Mims, T. Eckman, and C. Goldman, “Time-varying value of electric energy efficiency,” Jun....
  • A. Molina-Markham, P. Shenoy, K. Fu, E. Cecchet, and D. Irwin, “Private memoirs of a smart meter,” in Proceedings of...
  • Y. Wang et al.

    Review of smart meter data analytics: applications, methodologies, and challenges

    IEEE Trans. Smart Grid

    (2019)
  • Cited by (0)

    View full text