A Cross-Dimensional Attention Discriminating Masked Method for Building Energy Time-Series Data Imputation

Missing data in building energy time series is a pervasive issue, which leads to data format inconsistencies and hindrances in energy prediction and management. The most common approach to addressing missing data in building energy data is through data imputation. The critical challenge of data imputation is ensuring that the imputed data closely approximates the real values. This paper proposes a Cross-Dimensional Attention Discriminating Masked (CDADM) method, which modifies the self-attention model. The CDADM method captures essential global information about missing values through the cross-dimensional attention mechanism. Meanwhile, the discriminating attention mask mechanism enhances the ability to extract dependencies in long sequence data. Experimental results on typical building energy consumption data demonstrate that our proposed model is better than existing data imputation methods and shows the generality of the CDADM model.