With the access of multiple renewable resources to virtual power plants, hundreds of millions of power time series data are generated every day. A sparse learning-based power data compression and reconstruction processing method is designed in the study, which effectively solves the problems of low computational efficiency in the data processing centre of the virtual power plant and the waste of storage resources. According to the vector principal component analysis method, the power data are compressed. Then the data reconstruction network model is constructed based on sparse learning to achieve the reconstruction of power data. The experimental test results show that the median absolute errors of reconstruction of active and reactive power data are 4.05 MW and 0.885 Mvar, respectively, and the percentages of absolute errors are not more than 5%, which makes the reconstruction performance highly stable. The method achieves high-quality power data compression and highprecision reconstruction processing, which is of great significance for improving the computational efficiency of the virtual power plant data centre and accelerating the digital transformation of the power grid.