Study on data augmentation with physics-informed generative adversarial networks and the extrapolation performance of COP prediction for chillers

•PI-WGAN framework improves COP prediction under sparse data and distribution shift.•Vendi Score and Mahalanobis distance quantify data diversity and extrapolation risk.•Physics-informed samples enhance physical consistency and statistical reliability.•MAPE reduced from 41.2 % to 136.3 % to 4.97 %–1...

Full description

Saved in:
Bibliographic Details
Published in:Energy conversion and management Vol. 346; p. 120418
Main Authors: Wang, Zhengyang, Chen, Jun, Guo, Kexin, Xu, Bo, Chen, Zhenqian
Format: Journal Article
Language:English
Published: Elsevier Ltd 15.12.2025
Subjects:
COP prediction
Data augmentation
Generative adversarial network
Physics-informed neural network
Uncertainty quantification
COP prediction
Uncertainty quantification
Data augmentation
Physics-informed neural network
Generative adversarial network
ISSN:0196-8904
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•PI-WGAN framework improves COP prediction under sparse data and distribution shift.•Vendi Score and Mahalanobis distance quantify data diversity and extrapolation risk.•Physics-informed samples enhance physical consistency and statistical reliability.•MAPE reduced from 41.2 % to 136.3 % to 4.97 %–10.02 % in typical extrapolation tasks.•Explicit polynomial model enables control integration and lightweight deployment. Chiller systems are widely applied in buildings, industries, and data centers, where the coefficient of performance (COP) serves as a key indicator of energy efficiency. To address challenges of data sparsity and distributional shift in COP prediction, this study constructs 13 representative extrapolation tasks based on the ASHRAE‑1043 dataset and systematically evaluates the generalization performance of multivariate polynomial regression under varied data partitioning strategies. The Vendi Score is introduced to quantify data diversity, and variable contribution analysis is used to identify high-error regions. Furthermore, a dual extrapolation risk assessment is developed by combining Mahalanobis distance and predictive variance, enhancing model reliability evaluation. To improve performance in sparse and boundary regions, a Physics-Informed Wasserstein Generative Adversarial Network (PI-WGAN) framework is proposed. By embedding monotonicity constraints derived from chiller thermodynamics into the GAN training, PI-WGAN generates synthetic samples that are both statistically plausible and physically consistent. Experiments show that PI-WGAN significantly reduces extrapolation errors, lowering mean absolute percentage error (MAPE) from 136.3 % to 4.97 % in representative tasks. It outperforms Backpropagation (BP) neural networks, Physics-Informed Neural Networks (PINN), Variational Autoencoders (VAE), and Conditional Tabular GANs (CTGAN) in both accuracy and robustness. Moreover, the resulting explicit polynomial model provides structural transparency and can be easily deployed on industrial platforms for real-time control and optimization. Compared to conventional deep models, the approach offers better interpretability, computational efficiency, and control integration. In summary, PI-WGAN presents a reliable and efficient solution for energy performance modeling under small-sample and distribution-shift scenarios in industrial systems.
ISSN:0196-8904
DOI:10.1016/j.enconman.2025.120418