Evaluating performance of different generative adversarial networks 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)
- et al.
Automatic generation and simulation of urban building energy models based on city datasets for city-scale building retrofit analysis
Appl. Energy
(2017) - et al.
Development of prediction models for next-day building energy consumption and peak power demand using data mining techniques
Appl. Energy
(2014) - et al.
Tuning machine learning models for prediction of building energy loads
Sustain. Cities Soc.
(2019) - et al.
Data-driven Urban Energy Simulation (DUE-S): A framework for integrating engineering simulation and machine learning methods in a multi-scale urban energy modeling workflow
Appl. Energy
(2018) - et al.
Data-driven Urban Energy Simulation (DUE-S): Integrating machine learning into an urban building energy simulation workflow
Energy Procedia
(2017) - et al.
An integrated data-driven framework for urban energy use modeling (UEUM)
Appl. Energy
(2019) - et al.
DUE-B: Data-driven urban energy benchmarking of buildings using recursive partitioning and stochastic frontier analysis
Energy Build.
(2018) - et al.
Data driven parallel prediction of building energy consumption using generative adversarial nets
Energy Build.
(2019) - et al.
Generating realistic building electrical load profiles through the Generative Adversarial Network (GAN)
Energy Build.
(2020) - et al.
A novel semi-supervised data-driven method for chiller fault diagnosis with unlabeled data
Appl. Energy
(2021)
A semi-supervised approach to fault detection and diagnosis for building HVAC systems based on the modified generative adversarial network
Energy Build.
Comparing Generative Adversarial Networks architectures for electricity demand forecasting
Energy Build.
Energy Information Administration
Transactive control of commercial buildings for demand response
IEEE Trans. Power Syst.
Western wind and solar integration study phase 3–frequency response and transient stability
National Renewable Energy Lab. (NREL), Golden, CO (United States)
Energy plus: energy simulation program
ASHRAE J.
Virtual EPB
Calibrating building energy models using supercomputer trained machine learning agents
Concurrency and Computation: Practice and Experience
Development of new baseline models for US medium office buildings based on commercial buildings energy consumption survey data
Sci. Technol. Built Environ.
Cited by (16)
Filling time-series gaps using image techniques: Multidimensional context autoencoder approach for building energy data imputation
2024, Applied Thermal EngineeringA scenario framework for electricity grid using Generative Adversarial Networks
2023, Sustainable Energy, Grids and NetworksResearch on future trends of electricity consumption based on conditional generative adversarial network considering dual-carbon target
2024, IET Energy Systems Integration