From Signals to Semantics: A Multilabel Power Quality Disturbance Identification Framework Using Large Language Models

Accurate identification of power quality disturbances (PQDs) under the complex dynamic variations introduced by high renewable energy penetration remains a critical challenge. To address this issue, this article proposes a large language model-empowered power quality disturbance identification (LLME-PQDI) framework that converts voltage signals into semantics. To bridge voltage signals and large language model (LLM)-processable texts, a PQD semantic embedding module is developed. Within this process, a sample entropy-weighted patch contrastive learning method is introduced to strengthen feature extraction by focusing on uncertain informative signal segments. Furthermore, to enhance the LLM’s understanding of disturbance characteristics, a disturbance-guided prompt generation (DGPG) strategy is established, incorporating key time-frequency domain features of different PQD categories. This semantic conversion essentially facilitates the application of the powerful reasoning and generalization capabilities of LLMs for robust identification against distributional shifts in disturbance features. Experimental results on complex datasets with expanded parameter ranges and severe noise demonstrate that LLME-PQDI outperforms state-of-the-art approaches, achieving an exceptional accuracy of 98.05%, thereby validating the effectiveness of the proposed identification framework.