In the traditional power sector, offline data warehouses have problems such as insufficient real-time processing capabilities, low classification efficiency of common exception handling methods, and core issues. This study proposes a comprehensive solution based on the combination of anomaly detection algorithms and data warehouses. The specific work and contributions are as follows: The differences among various feature selection methods were discussed and regarded as a prerequisite step for the anomaly detection algorithm to improve the model training speed. Based on this, an improved isolated forest algorithm is proposed. This algorithm utilizes binary Particle Swarm Optimization (BPSO) technology to solve the high redundancy problem of the base classifier in the original isolated forest algorithm. Aiming at the problem of dynamic threshold setting for abnormal scores, a sliding window adaptive threshold determination method based on the K-Means ++ algorithm is proposed.