Data Augmentation of Power Digital Twin Based on Visual Evaluation

When the historical data sample is insufficient, it is necessary to use algorithms to expand the data space of the power digital twin system. In this paper, the visual evaluation method based on deep neural network and few-shot learning are integrated to realize the multi-dimensional data augment of data space under the condition of limited historical samples. Firstly, multi-type simulation samples are generated based on Wasserstein GAN. Then the real and simulated samples are mapped to the low-dimensional visualization space by deep auto-code model. Through the evaluation in the visual space, the samples can be verified, and based on the results, the subsequent simulation samples of specific scenes can be generated, and the diversified dimension of data space can be realized. The proposed method was verified by a case study based on the measured data of a real power system.