Multi-agent collaborative semantic recognition and real-time event generation for power grid alarm monitoring

Massive alarm data from grid monitoring systems is essential for secure power dispatch operation. Aggregating discrete alarm signals into structured events is vital to enhance grid monitoring intelligence. Nevertheless, existing event rule bases adopt domain-customized complex grammar with diverse symbolic expressions, making traditional string matching methods unable to achieve accurate semantic parsing. This paper proposes a multi-agent collaboration-based method for semantic recognition and real-time generation of power grid monitoring alarm events. The method integrates large language models (LLMs), knowledge graph (KG) representation learning, and multi-agent coordination to realize high-precision semantic understanding and real-time intelligent generation of alarm events. Specifically, three original contributions are made: (1) a three-layer standardized preprocessing pipeline combined with instruction tuning and retrieval-augmented generation (RAG) that bridges the semantic gap between domain-specific rule grammar and LLMs; (2) a dual-drive fusion recognition model integrating GraphSAGE-based knowledge graph embedding and BiLSTM temporal encoding with an attention fusion mechanism; and (3) a hierarchical multi-agent collaborative architecture with explicit communication protocols, conflict resolution strategies, and a SEEN (Semantic Event Exception Notifier) algorithm for unknown event detection. Experimental results on 356,950 real alarm samples from a regional power grid demonstrate that the proposed method achieves an F1-score of 96.9% and a false alarm rate of 4.2%, outperforming conventional rule matching methods by 25.3% points under signal-loss scenarios. Ablation studies confirm that each module contributes independently and significantly to the overall performance.