Enhanced Named Entity Recognition in Power Grid Operations: A Span-based Approach for Chinese Dispatch Communications

Named Entity Recognition (NER) in power grid dispatch operations is critical for operational safety and system reliability. This domain faces unique challenges including domain-specific terminology, nested entity structures, and Chinese language processing complexities. This work presents GP-NER, a novel span-based neural framework designed to extract structured entities from Chinese power grid voice dispatch instructions, including domain-specific terminology and nested entity structures. Our architecture combines Chinese-RoBERTa with whole-word masking, BiLSTM sequential modeling, and dual feed-forward networks to generate specialized representations for entity classification and span encoding. GP-NER integrates OpenAI’s Whisper for speech-to-text conversion with downstream entity extraction across nine power grid-specific categories. The core innovation lies in our similarity-based span-type matching mechanism that models semantic relationships between entity types and candidate spans, enabling robust handling of nested and overlapping entities prevalent in technical dispatch communications. Comprehensive 10-fold cross-validation on 2,000 authentic dispatch recordings demonstrates superior performance with 94.93% precision, 96.10% recall, and 95.51% F1-score, significantly outperforming competitive baselines including recent span-based methods. Statistical significance testing confirms these improvements (all p < 0.05). Ablation studies validate individual component contributions, while evaluation on public Chinese NER benchmarks demonstrates generalizability beyond the power grid domain. The framework shows consistent effectiveness across all entity categories, with F1-scores ranging from 94.85% to 95.99%. GP-NER provides an effective solution for domain-specific entity extraction in power grid environments, advancing intelligent grid management through enhanced automated processing and operational safety.