Abstract: Aiming at the problems of dense terms, semantic ambiguity and high response delay requirements in scheduling speech, we propose a semantic understanding method based on lightweight graph attention mechanism. The method fuses the pre-trained language model and the power domain ontology knowledge to construct a semantic parsing framework with structure-aware ability. By optimizing the structure of the bidirectional graph attention module, we introduce parameter sharing and edge weight simplification strategies to effectively compress the complexity of the model, and disambiguation of the "device-operation-parameter" semantic structures in the scheduling instructions. A hybrid generation decoder is designed, which combines the generation, selection and copy three-way gating strategy to effectively deal with the structural generation of unknown words and complex terms. Experimental results show that the proposed method achieves 93.5% intent recognition accuracy and 91.8% slot-filling F1 value on the Xinjiang power grid dispatching corpus, and still maintains 85.7% generalization performance in cross-domain tests. The system response delay is as low as 68 ms, which is significantly better than the mainstream baseline methods.