A Multi-view Learning-Based Approach for Handling Missing Values in Building Energy Data

Building energy data typically suffer from missing data problems which may seriously affect the quality of subsequent data analysis. Existing studies mainly focused on proposing customized missing value processing methods for individual buildings. To effectively utilize existing data resources in the building field and improve the versatility of missing value imputation methods for energy consumption data, this study puts forward solutions based on the concept of multi-view learning. The solutions are developed considering four views, i.e., local temporal view, local similarity view, global temporal view and global similarity view. Various machine learning methods have been adopted for each view learning respectively, including moving average, collaborative filtering, long-short term memory network and support vector regression. To effectively capture the global and local data variations and improve the overall accuracy, the ensemble learning method is then applied to incorporate results from multi-view learning. Data experiments have been designed to test the effectiveness in handling random and continuous missing data patterns, based on which the results on multiple school buildings are reported and discussed. The research outcomes are helpful for enhancing the accuracy and generalizability of missing value imputation methods for building energy data analyses.