Abstract: To improve the automation and standardization of power grid dispatching instruction generation, this paper proposes a method integrating BERT scene classification with Seq2Seq text generation. A four-stage closed-loop workflow covering state perception, scene recognition, instruction generation, and compliance verification is designed for day-ahead distribution network dispatching. Comparative experiments based on 8640 historical dispatching records show that the proposed method significantly outperforms both the manual and template-based approaches in compliance rate, terminology accuracy, operation sequence correctness, and generation efficiency, providing a feasible reference for intelligent upgrading of dispatching operations.