The missing data in electricity consumption can affect the performance of electricity theft detection models

while traditional interpolation methods is difficult to retain the key features and limit the performance of the model. This paper proposes a two-stage self-attention weight fusion (TS-SAWF) model for missing data imputation in electricity theft detection. This model utilizes a two-stage convolution-enhanced diagonal masked multi-head self-attention module to capture the correlations and features between the data. It further integrates attention weights from both stages to improve imputation accuracy. Experimental results on two real-world electricity meter datasets from the State Grid Corporation of China demonstrate that

compared to the comparison models

the TS-SAWF model achieves higher imputation performance under varying random missing rates and different numbers of consecutive missing days. Moreover

the data imputed by the TS-SAWF model effectively enhances the performance of the electricity theft detection model

further demonstrating its effectiveness.