Abstract: Taking a categorical principal components analysis data processing module based on optimal scale quantization and an optimized Transformer time series forecast module as main module, a satellite power consumption prediction method was proposed. Aiming at the high redundancy problem of satellite engineering data, a satellite high-dimensional data processing model based on Hurst index analysis, grey relational analysis and categorical principal component analysis (CATPCA) was established to effectively extract hundred-dimensional time series data and reconstruct the input data. In addition, the adversarial learning network architecture was used to establish a satellite power prediction model of multi-learning Transformer. The model was designed to comprehensively consider various affecting factors on satellite energy consumption and time series data dependencies, and to complete high-precision satellite power consumption time series prediction in a short period of time. In the experiment part, adopting the real operation data of satellite and comprehensively considering various factors that affect satellite energy consumption, the fitting accuracy of proposed method can reach up to 94% with 12h prediction, which is higher than that of BP neural network and long short-term memory network (LSTM). The results show that the method can effectively overcome the problems of redundancy, lack and dirty data of conventional engineering data, solve the deficiency that conventional time series prediction needs to rely on long-term data, effectively complete the high-precision prediction of satellite energy consumption in a short time. This provides reliable support for satellite on-orbit mission planning, satellite on-orbit health management and other follow-up tasks, and assists decision-making.