Anomaly Detection Method for Power Dispatch Streaming Data Based on Adaptive Isolation Forest and Self-Su-pervised Learning

Tenglong Xie^*Bo LiYingkai ZhenHao WeiKunhuan XuJingyin Huang

• Information Center of Guangdong Power Grid Co., Ltd. GuangZhou 510600 China, guangzhou, China

To address the issues of concept drift and scarcity of anomaly samples in real-time anomaly detection under the massive streaming data environment of power dispatching and control systems. We propose a streaming data anomaly detection method that integrates adaptive isolation forests and self-supervised learning. First, by analyzing the inherent relationships between system services, processes, and resource usage, we construct a business-based model for anomaly detection. Furthermore, we propose an improved isolation forest algorithm based on a sub-forest progressive update mechanism. By selectively eliminating sub-detectors with large anomaly rate deviations and dynamically adding new detectors, we overcome the model performance degradation caused by traditional random update strategies and significantly improve the detection algorithm's adaptability to concept drift and overall stability. To further explore the temporal features in streaming data, we introduce a self-supervised learning framework based on the GPT architecture. We design state memory units to encode historical data patterns and reduce data redundancy through a sampling strategy based on distance metrics, effectively enhancing the ability to perceive hidden anomalies. Experiments on a real power dispatching process resource dataset show that the proposed method significantly out-performs the traditional streaming data isolation forest algorithm in key indicators such as AUC value, with the highest improvement reaching 39.12%. Ablation experiments verify the effectiveness of each module, providing reliable technical support for the safe and stable operation of the power dispatching system.