.png){kind=icon}
.png){kind=icon}
Return
A missing data imputation method for chillers based on denoising diffusion probabilistic model
.png){kind=icon}
PRE
AI
J
X
Y
DOI：10.1016/j.enbuild.2026.118067.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Chillers are frequently affected by harsh operating conditions and noise interference, resulting in substantial missing values in the raw data collected by sensors. These missing values can significantly undermine the reliability of energy efficiency analysis and fault diagnosis. Moreover, chiller data are typically characterized by high dimensionality, strong multivariate coupling, and nonlinear as well as non-stationary temporal dynamics, which pose considerable challenges for accurate imputation of multivariate time series. To address these issues, this paper proposes a novel time series imputation framework based on a denoising diffusion probabilistic model (DDPM). Multivariate time series are first transformed into two-dimensional grayscale representations. Sparse observed points selected by control point sampling are then used as conditional information, and an improved ConvNeXt-V2 module is integrated into a U-Net denoising network. During reverse diffusion, the control point observations and missing-value masks guide the reconstruction of missing values. To demonstrate the effectiveness of the proposed method, extensive experiments are conducted on two datasets under two missing value patterns, with missing rates ranging from 10% to 90%, and a large number of comparative experiments are performed. For all missing data scenarios, the average RMSE, MAE, and NMAD values of the method proposed in this paper were reduced by 15.8%, 20.9%, and 20.2%, respectively, compared to the best-performing baseline model. The experimental results validate that the proposed approach achieves superior imputation accuracy and generalization performance.
Keywords:
Chillers
Missing data imputation
DDPM
Multivariate time series
Control point sampling
Journal
IF:
7.1
Papers:
1.5W
Citations:
6.8W
Organization
Cited Papers