A text-table fusion model for electricity trading and cross-modal knowledge fusion

To address the challenges of multi-modal data and complex rules in the electricity trading of new-type power systems, this paper proposes a text-table fusion large model for electricity trading (Power-FusionLLM). Current electricity trading decision support systems mostly rely on single-modal data, making it difficult to effectively associate unstructured policy text with structured trading tables, resulting in a “semantic gap”. Power-FusionLLM achieves domain adaptation by constructing a proprietary instruction tuning dataset for the power domain. It designs a strategy-guided Policy-Gated Cross-Attention (PGCA) mechanism, enabling the model to understand how text rules constrain or influence tabular data. To verify model performance, this paper constructs a high-fidelity simulation and desensitized historical trading fusion dataset covering three typical tasks: market joint QA, trading strategy generation, and compliance detection. Experimental results show that Power-FusionLLM significantly outperforms benchmark models in all tasks, achieving an accuracy of 89.5% in market QA and reducing the Numerical Calculation Error Rate (CER-N) to 3.8%. Ablation studies and case analyses further confirm the effectiveness of the cross-modal fusion mechanism. This study provides a new solution path for the deep application of large language models in vertical domains.