The digital twin grid generates vast amounts of multimodal data daily, comprising text and images. Analyzing the sentiment within this massive dataset is crucial for understanding power grid scheduling and decision-making. This paper focuses on multimodal sentiment analysis, acknowledging the complementary nature of information in text and images. To effectively capture emotional characteristics and intermodal interactions, a multimodal digital twin power sentiment analysis model is proposed based on the attention mechanism. Two attention-based unimodal feature extraction models are introduced for text and image sentiment features. A tensor fusion strategy is employed to obtain joint multimodal feature representation for sentiment classification. Attention neural networks highlight key regions and words containing emotional information in images, enhancing the precision of unimodal feature representation. Principal component analysis eliminates redundant information in joint features, and support vector machines categorize sentiment in multimodal data. The proposed approach offers a comprehensive and precise understanding of power users’ emotions in the digital twin grid, contributing to improved decision-making in power grid management.