Accurate multivariate prediction of spacecraft power systems supports energy management and mission continuity. Prediction methods based on mechanistic models and data-driven approaches have been applied to power system operation analysis with some success. However, reducing model dependence while maintaining multivariate prediction accuracy under changing conditions and longer prediction horizons remains challenging. To address these challenges, this paper proposes a residual-compensated deep Koopman network for multivariate multi-step prediction through latent state evolution modeling and historical error correction. First, a deep Koopman network is constructed to update latent states through a shared linear operator and generate base forecasts using a decoder. Second, latent consistency and multi-step prediction losses are combined to jointly learn state representations relevant to the prediction task and their evolution relationships. Finally, a historical residual compensation module is developed, in which residuals and recent operating states are jointly fed into a GRU to generate corrections for each forecast step. Experiments under normal and abnormal operating conditions show that the proposed method reduces MSE by approximately 13.20% compared with GRU, the best-performing baseline, improving multivariate prediction accuracy. The method can provide predictive support for assessing spacecraft power system operating states and informing subsequent scheduling decisions.

multivariate prediction,spacecraft power systems,residual compensated,Koopman Network