Feasibility Analysis of Anomaly Detection in Power System Data Markup Language Based on Lightweight DeepSeek

Anomaly detection for data markup language in power systems is a critical technology for ensuring reliable data transmission and processing. To address the challenge of complex syntactic and semantic anomalies, this paper proposes a lightweight DeepSeek model-based anomaly detection framework, and investigates its application feasibility in power systems. The proposed approach employs Word2Vec and K-means algorithms to extract representative E-language texts from file names. By combining the contextual learning capabilities of large models, the approach enables real-time online detection of E-text with only a small number of samples. This work focuses on identifying syntactic anomalies such as missing fields, and semantic anomalies including out-of-range values and type errors. Comparative experiments have been conducted among three distilled versions of DeepSeek-R1, where the DeepSeek-R1-Distill-QWEN-14B model achieves the best performance, with an F1 score of 86.6%, while attaining a sound balance between performance and resource efficiency. The results demonstrate the framework's potential for anomaly detection in power system data markup language, and offer new insights for its development.