Abstract: Accurately measuring data are essential for constructing high-precision data-driven models. However, the actual operational data of a power station are subject to a number of influences, such as measurement, communication, storage and other processes, inevitably resulting in partial data loss, which can subsequently impact the effectiveness of modeling based on the operational data. To address this issue, a missing data imputation method based on reconstructed variational autoencoder was proposed. This method constructed reconstruction indexes based on the variational self-coding model and utilized the Steffensen method to accelerate the iterative updating of single value or multiple missing values. This approach ensures computational efficiency while maintaining interpolation accuracy, allowing for the efficient and accurate interpolation of single missing datum or multiple missing data. Mathematical calculations and engineering examples were employed to analyze the results. When the number of missing parameters is 1, 2, 3 and 6, the mean coefficients of determination (R²) of the reconstruction-based variational autoencoder interpolation results are 0.998 0, 0.998 6, 0.999 2 and 0.996 2, respectively. These values demonstrate minimal deviation from the actual values, and the interpolation effect is evidently superior to that of the missForest and generative adversarial imputation net (GAIN) methods.