Modeling of source–grid–load–storage synergy in microgrids using a multimodal graph transformer

This study proposes a multimodal graph Transformer-based modeling method for source–grid–load–storage synergy in microgrids. The method constructs a global directed graph and employs a graph convolutional network to achieve spatiotemporal alignment and global feature extraction from multimodal data, effectively mitigating the loss of inter-modal correlations inherent in conventional models. Furthermore, a dynamic cross-modal attention mechanism is introduced to adaptively adjust attention weights according to the real-time coupling strength among different modalities, thereby enhancing the model’s adaptability to complex operating conditions. A hierarchical multimodal feature fusion and weighting strategy is also adopted to exploit the complementarity among modalities and improve overall performance. Experimental results on the Solar Radiation Research Laboratory dataset demonstrate that the proposed model achieves an average F 1-score improvement of 2.75% over the baseline Transformer and 2.17% over the QL-Transformer. Compared with the multimodal cross-attention Transformer ( F 1 = 0.852).