In response to the issues of neglecting prior information and low computational efficiency of the traditional Singular Value Thresholding （SVT） algorithm in power load data recovery， a novel improved SVT algorithm based on phase space reconstruction and adaptive variable step length is proposed. To address the problem of neglecting prior information in traditional SVT， a phase space reconstruction algorithm is introduced to map the original missing data into a high-dimensional space， leveraging data correlation and structural features as prior knowledge for subsequent data recovery algorithms. By combining logarithmic and sigmoid functions to construct the variable step length base function， and utilizing geometric progression to enhance the initial step length， an adaptive variable step length SVT algorithm is built to overcome the low computational efficiency issue of traditional SVT in large-scale data scenarios. Comparative experimental analysis is conducted using multiple publicly available power load datasets and various commonly used power load data recovery algorithms. The results demonstrate that the improved SVT algorithm achieves better data recovery performance， with enhanced convergence speed， accuracy， and stability， showcasing strong engineering practicality.