Generating realistic building electrical load profiles through the Generative Adversarial Network (GAN)
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)
- et al.
A method of formulating energy load profile for domestic buildings in the UK - ScienceDirect
Energy Build.
(2005) - et al.
Linking energy-cyber-physical systems with occupancy prediction and interpretation through WiFi probe-based ensemble classification
Appl. Energy
(2019) - et al.
Data fusion in predicting internal heat gains for office buildings through a deep learning approach
Appl. Energy
(2019) - et al.
Predicted vs. actual energy performance of non-domestic buildings: Using post-occupancy evaluation data to reduce the performance gap
Appl. Energy
(2012) - et al.
Residential heating energy consumption modeling through a bottom-up approach for China’s Hot Summer-Cold Winter climatic region
Energy Build.
(2015) Price sensitivity of residential energy consumption in Norway
Energy Econ.
(1999)- et al.
Comparison of neural network, conditional demand analysis, and engineering approaches for modeling end-use energy consumption in the residential sector
Appl. Energy
(2008) - et al.
Modeling of the space and domestic hot-water heating energy-consumption in the residential sector using neural networks
Appl. Energy
(2004) - et al.
Modeling of end-use energy consumption in the residential sector: A review of modeling techniques
Renew. Sustain. Energy Rev.
(2009) - et al.
The Building Data Genome Project: An open, public data set from non-residential building electrical meters
Energy Procedia
(2017)