Elsevier

Energy and Buildings

Volume 269, 15 August 2022, 112247
Energy and Buildings

Evaluating performance of different generative adversarial networks for large-scale building power demand prediction

https://doi.org/10.1016/j.enbuild.2022.112247Get rights and content

Highlights

  • Develop a new strategy to evaluate GAN accuracy for building power demand prediction.

  • Identify and evaluate five GANs for predicting large-scale building power demand.

  • Recommend Original GAN and cGAN for large-scale building power demand prediction.

Abstract

As an unsupervised-learning data-driven model, Generative Adversarial Networks (GANs) have recently attracted a lot of attention for various applications. There is potential to apply GANs for large-scale building power demand prediction, which is needed for power grid operation. However, there are many GAN variations and it is unclear which GAN is suitable for this application. To answer this question, this paper identifies five promising GANs (Original GAN, cGAN, SGAN, InfoGAN, and ACGAN) and evaluates their performance for predicting building power demand at a large scale. Physics-based building energy models are developed to generate training and reference data. A new evaluation indicator that combines accuracy and reproducibility is proposed to evaluate the performance of different GANs in predicting building power demand. The results show that SGAN and InfoGAN are not suitable because they cannot control the number of generated building samples for different building types. The prediction performance among the Original GAN, cGAN, and ACGAN can vary depending on training sample sizes and number of building types. If the training sample size is sufficiently large, Original GAN and cGAN can predict building power demand more accurately than ACGAN with the same number of samples. If training samples are limited, Original GAN provides better accuracy than cGAN and ACGAN. When the number of building types increase, the prediction accuracy increases for cGAN, decreases for ACGAN, and remains the same for Original GAN. As a result, cGAN and Original GAN are recommended for large-scale building power demand prediction.

Introduction

Rapid deployment of electrification to address goals of reducing carbon emissions will lead to significant challenges to the power grid. Buildings show great potential for addressing these challenges, since they consume approximately 75% of electricity in the U.S [1] and can be flexible demand-side resources to support the optimal and stable operation of the power grid [2], [3]. To investigate these new challenges, the first step is to precisely predict building power demand at a large-scale. Physics-based simulation methods are often used for large-scale building power prediction. These methods rely on building energy simulation tools, such as DOE-2 [4] and EnergyPlus™ [5]. Urban scale energy modeling and simulation platforms, such as Virtual EPB [6], CitySim [7], and URBANopt [8], were all developed based on building energy modeling tools. While physics-based methods have been proven to be useful for predicting building power demand at a large scale, they have some limitations. This includes requiring expert knowledge and complex workflows to create and calibrate models [9], [10], as well as significant computational time to run hundreds of thousands of detailed building energy simulations for large-scale building power demand prediction.

Data-driven methods are another option for predicting building power demand, which can address the limitations of physics-based methods. These methods utilize historical data to predict future power demand [11], [12], [13]. Recently, some data-driven urban-scale modeling frameworks were developed to predict large-scale building power demand [14], [15], [16], [17]. For example, Yang et al. [17] developed a data-driven urban-scale modeling framework for energy benchmarking of buildings based on recursive partitioning and stochastic frontier analysis. Over 10,000 buildings were involved in this study. The results indicated that this data-driven framework is robust and able to generate accurate energy benchmarking. However, one common limitation of data-driven methods is the requirement of sufficient historical data to provide accurate predictions. Generative adversarial networks (GANs) have become a popular unsupervised-learning data-driven method for various applications that address this limitation. A GAN model is not only able to capture the features of provided data, but also able to keep the individualities of each sample [18]. Torres et al. [19] and Tian et al. [20] pointed out that a GAN model is potentially able to accurately predict time-series building power demand with limited information. This makes GANs a potentially good option for fast and accurate large-scale building power demand prediction with a limited need for historical data.

However, the original GAN model has limitations for more complex studies. For example, the original GAN model is designed to generate samples for a single class, while large-scale power demand prediction involves multiple classes (e.g., building types). To solve this issue, some researchers conducted clustering to classify building samples and then, trained a GAN model for each class [21]. Recently, hundreds of variations of GANs have been proposed to address the limitations of the original GAN model [22], including models with the ability to generate samples for multiple classes. Besides first-clustering-then-GAN approach, these variations of GANs are also potential candidates for multiple-class study. It remains unclear, however, which of these GANs are suitable for building power demand prediction at a large scale.

To answer this question, this paper identifies five promising GANs (Original GAN, cGAN, SGAN, InfoGAN, and ACGAN) and evaluates their performance for predicting building power demand at a large-scale. Physics-based building energy models are developed to generate training and reference data. In addition, a new evaluation indicator is proposed to evaluate the performance of different GANs to predict building power demand. Our scientific contributions include: 1) systematically evaluating the performance of different GAN models for large-scale building power demand prediction, 2) developing a new evaluation indicator to combine the impacts of prediction accuracy and reproducibility, and 3) identifying two GAN models that are not suitable for large-scale building power demand prediction with current methods, as well as one improved GAN model that performs well for several building types and limited training samples.

The rest of this paper is organized as follows: Section 2 selects candidate GANs used in this study; Section 3 proposes the evaluation strategy to determine the suitable GANs for large-scale building power demand prediction; Section 4 prepares training data and reference data using physics-based building energy models; Section 5 displays the evaluation results and conducts discussion based on the results; and finally, Section 6 provides conclusions.

Section snippets

GANs used in the study

This section introduces variations of GAN and selects a few GANs potentially suitable for predicting large scale building power demand to be further investigated in this study. SubSection 2.1 reviews different GANs and selects five GANs: Original GAN, cGAN, SGAN, InfoGAN, and ACGAN; SubSection 2.2 introduces the Original GAN; SubSection 2.3 introduces the other four selected variations of GAN created based on the Original GAN.

Evaluation strategy

This section introduces a strategy to evaluate the five studied GANs for large-scale building power demand prediction in two aspects: suitability and sensitivity. SubSection 3.1 investigates the evaluation strategy for the model suitability by setting a series of application requirements. SubSection 3.2 introduces the evaluation strategy for the sensitivity analysis. SubSection 3.2.1 introduces the evaluation indicator used in the sensitivity analysis while SubSection 3.2.2 introduces the

Building power data preparation

This section describes the preparation of building power data. This study mainly aims to evaluate the performance of selected GANs for large-scale building power demand. To simplify the study and avoid potential noise in the data, this paper uses physics-based models to generate building power data. We consider different building types, and uncertainties of building characteristics and operation in the building set. The seed models are from DOE’s Commercial Prototype Building Models [37] and

Evaluation results and discussion

This section analyzes the evaluation results by using the strategy introduced in Section 3. We then discuss the results and recommend the GANs for large-scale building power demand prediction. SubSection 5.1 introduces the model settings for all five studied GANs; SubSection 5.2 analyzes the model suitability for large-scale building power demand prediction; Section 5.3 conducts sensitivity analysis to evaluate the impacts of training size and number of building types to the evaluation

Conclusion

This paper develops a novel strategy to evaluate the performance of different GANs for predicting building power demand at a large scale. Five promising GANs (Original GAN, cGAN, SGAN, InfoGAN, and ACGAN) are identified and evaluated using this strategy. Based on the result, Original GAN and cGAN are recommended, while SGAN, InfoGAN, and ACGAN are excluded from the candidate list for research about large-scale building power demand prediction. With a limited number of training samples, Original

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 (45)

  • B. Li et al.

    A semi-supervised approach to fault detection and diagnosis for building HVAC systems based on the modified generative adversarial network

    Energy Build.

    (2021)
  • N.M.M. Bendaoud et al.

    Comparing Generative Adversarial Networks architectures for electricity demand forecasting

    Energy Build.

    (2021)
  • B.P. Center et al.

    Energy Information Administration

    (2020)
  • H. Hao et al.

    Transactive control of commercial buildings for demand response

    IEEE Trans. Power Syst.

    (2016)
  • N.W. Miller et al.

    Western wind and solar integration study phase 3–frequency response and transient stability

    National Renewable Energy Lab. (NREL), Golden, CO (United States)

    (2014)
  • F. Winkelmann, B. Birdsall, W. Buhl, K. Ellington, A. Erdem, J. Hirsch, S. Gates, DOE-2 supplement: version 2.1 E,...
  • D.B. Crawley et al.

    Energy plus: energy simulation program

    ASHRAE J.

    (2000)
  • J.A. Ingraham et al.

    Virtual EPB

    (2018)
  • B. Polly, C. Kutscher, D. Macumber, M. Schott, S. Pless, B. Livingood, O. Van Geet, From zero energy buildings to zero...
  • J. Sanyal et al.

    Calibrating building energy models using supercomputer trained machine learning agents

    Concurrency and Computation: Practice and Experience

    (2014)
  • Y. Ye et al.

    Development of new baseline models for US medium office buildings based on commercial buildings energy consumption survey data

    Sci. Technol. Built Environ.

    (2020)
  • V. Cherkassky, S.R. Chowdhury, V. Landenberger, S. Tewari, P. Bursch, Prediction of electric power consumption for...
  • Cited by (16)

    View all citing articles on Scopus
    View full text