The significant increase in data volume generated by smart grid devices presents challenges for traditional power grid communication systems. Semantic communication offers a promising approach by extracting relevant features from the source information. However, existing large-scale model-based semantic communication systems often overlook the common MIMO channel conditions prevalent in power grid environments. To address this issue, we propose a Large Scale Model-aided Digital MIMO semantic communication system in Smart Grid (LSM-MIMO-SCSG), where the transmitter utilizes the LSM as the semantic encoder to extract domain-specific knowledge for smart grid applications and a semantic decoder is performed to accomplish smart grid task. Besides, a robust compression mechanism using vector quantization through the MIMO channel is proposed to quantize the extracted features into indices with a pretrained codebook. To optimize the semantic encoder/decoder and codebook design, a two-stage training strategy on a smart grid fault classification dataset is proposed. Experimental results show that our proposed LSM-MIMO-SCSG can achieve 34.92% classification accuracy and reduce 99.56% transmission symbols at most compared with the traditional methods.