ALBERT

Description: The integrity of data is an essential basis for analyzing power system operating status based on data. Improper handling of measurement sampling, information transmission, and data storage can lead to data loss, thus destroying the data integrity and hindering data mining. Traditional data imputation methods are suitable for low-latitude, low-missing-rate scenarios. In high-latitude, high-missing-rate scenarios, the applicability of traditional methods is in doubt. This paper proposes a reconstruction method for missing data in power system measurement based on LSGAN (Least Squares Generative Adversarial Networks). The method is designed to train in an unsupervized learning mode, enabling the neural network to automatically learn measurement data, power distribution patterns, and other complex correlations that are difficult to model explicitly. It then optimizes the generator parameters using the constraint relations implied by true sample data, enabling the trained Generator to generate highly accurate data to reconstruct the missing data. The proposed approach is entirely data-driven and does not involve mechanistic modeling. It can still reconstruct the missing data in the case of high latitude and high loss rate. We test the effectiveness of the proposed method by comparing three other GAN derivation methods in our experiments. The experimental results show that the proposed method is feasible and effective, and the accuracy of the reconstructed data is higher while taking into account the computational efficiency.