To address the problems of traditional anomalous power usage detection in the face of the curse of dimensionality in high-dimensional data, and the impact of irrelevant features on anomaly detection, resulting in low detection accuracy. An isolated forest detection algorithm based on unsupervised density subspace selection was proposed. First, an effective density-based compact data representation method was proposed to improve the efficiency of the subspace selection strategy. Then, the minimum redundancy-maximum correlation-density criterion (mRMRD) was applied to select the relevant mutual information-based subspaces. Finally, isolated trees are constructed and isolated forests were integrated in the correlated subspaces to achieve the detection of abnormal electricity consumption data. Through experimental analysis, the proposed method improves the accuracy, area under the ROC curve (area under curve, AUC), and F₁ indexes compared with the traditional detection algorithm, which effectively improves the abnormal electricity usage detection. Meanwhile, the sensitivity analysis also verifies the effectiveness of the unsupervised density subspace isolated forest detection algorithm.