ABSTRACT The digital management of distribution transformer areas in smart grids faces several core challenges. These include difficulties in fusing multi‐source heterogeneous data, high costs associated with real‐time monitoring, and insufficient accuracy of digital twins. Existing spatio‐temporal graph neural networks rely on complete data inputs and exhibit sensitivity to partial data loss caused by communication interruptions. While compressed sensing techniques can reduce sampling rates, their measurement and reconstruction processes are optimized independently. This approach overlooks the inherent spatio‐temporal correlations within grid topologies. To address these issues, a spatio‐temporal graph convolutional compressed sensing model is proposed. This model decouples spatio‐temporal features using dynamic graph convolution. It designs a topology‐aware measurement matrix to optimize sparse sampling. Furthermore, it incorporates physics‐constrained adversarial training to enhance the fidelity of digital twins. A learnable adjacency matrix mechanism is employed, utilizing node attention to dynamically adjust topological weights. This mitigates the over‐smoothing issue common in traditional graph convolutions during equipment switching operations. Additionally, a multimodal joint embedding layer is introduced. This layer maps unstructured data—such as infrared images—into graph node features, enabling cross‐modal correlation modeling. Experimental results demonstrate the model's effectiveness. At a 15% sampling rate, the proposed model reduces reconstruction error by 37.2% compared to baseline spatio‐temporal graph neural networks. In digital twin simulations, the voltage waveform correlation coefficient reaches 0.98. Furthermore, photovoltaic output prediction error is reduced by 14.6%. This framework provides a high‐accuracy, low‐cost digital twin solution for state estimation and fault early warning in distribution transformer areas.
Shi et al. (Wed,) studied this question.