Abstract: Real time monitoring and anomaly detection of power system line losses are core requirements for improving the economic operation efficiency of the power grid. However, traditional methods are limited by problems such as large noise interference from single source data, low anomaly detection accuracy, and response delay. Based on this, this article focuses on the dynamic fusion and abnormal feature mining of multi-source heterogeneous data, and proposes a line loss detection system that combines Kalman filtering and random forest. The Kalman filter is used to achieve real-time noise reduction and fusion of multi-dimensional data, and the random forest ensemble learning mechanism is combined to improve the ability to identify anomalies in complex scenes. Experiments have shown that this method achieves the optimal balance between detection accuracy (95.6%) and processing efficiency (5.3 s), and can reduce line loss rate in aging scenarios by up to 20%, providing a dynamic optimization solution for precise loss reduction in power systems.