A Filling Method Based on K-Singular Value Decomposition (K-SVD) for Missing and Abnormal Energy Consumption Data of Buildings

Abstract

Massive data can be collected from meters to analyze the energy use behavior and detect the operation problems of buildings. However, missing and abnormal data often occur for the raw data. Effective data filling and smoothing methods are required to improve data quality before conducting the analysis. This paper introduces a data filling method based on K-SVD. The complete dictionary is trained and then utilized to reconstruct incomplete samples to fill the missing or abnormal data. The impacts of the dictionary size, the data missing continuity, and the sample size on the performance of the proposed method are studied. The results show that a smaller dictionary size is recommended considering the computational complexity and accuracy. The K-SVD method outperforms traditional methods, showing a reduction in the MAPE and CVRMSE by 3.8–5.4% and 6.7–87.8%. The proposed K-SVD filling method performs better for non-consecutive missing data, with an improvement in the MAPE and CVRMSE by 0.1–4% and 5.1–6.7%. Smaller training samples are recommended. The method proposed in this study would provide an effective solution for data preprocessing in building and energy systems.

Keywords

• K-SVD
• data filling
• missing data
• data pre-processing
• building energy