Knowledge Distillation Method for Compressing Large Language Model of Power Risk Identification and Improving Deployment Efficiency

Abstract

Large language models demonstrate high precision in power-system risk identification; however, their massive parameter counts and high resource consumption hinder real-time deployment on resource-constrained edge devices used in electromagnetic sensing systems, wearable monitoring terminals, and compact power-monitoring devices. Achieving low-latency processing in distributed sensing structures and edge-based electromagnetic monitoring devices requires a significant reduction in computational overhead to ensure immediate detection of electrical hazards, abnormal equipment states, and potential power-system risks. This paper proposes a lightweight compression method based on hierarchical supervised knowledge distillation. Experiments show that, after applying the proposed knowledge distillation method, the inference latency is reduced from 235 ms to a minimum of 26 ms, which is better than DistilBERT’s 35 ms. The number of student model parameters is reduced to 4.3% of the teacher model, namely 14.5M versus 340M, while the classification accuracy reaches 89.4%, close to the teacher model’s 92.7%. The F1 score for the equipment failure category reaches 90.3%, verifying the efficiency and practicality of the lightweight model in resource-constrained scenarios involving power-risk identification, electromagnetic sensing, and edge-based intelligent monitoring.