Driven by the dual carbon goals, the widespread adoption of electric vehicles (EVs) has been accelerated by the demand for low-carbon and green transportation. To enhance the accuracy of electric vehicle charging load forecasting, supporting stable grid operation and optimized scheduling of charging infrastructure, a decomposition ensemble learning prediction (DEP) framework based on two-layer optimization and the unscented Kalman filter (UKF) was proposed for EV charging load forecasting. Firstly, a BKA-VMD decomposition model was constructed. The black-winged kite algorithm (BKA) was employed to adaptively optimize the parameters of variational mode decomposition (VMD) to enhance decomposition stability and modal feature extraction capability. Secondly, the improved polar lights optimization (IPLO) algorithm was applied to optimize the hyperparameters of the forecasting models. The UKF-Transformer-LSTM, UKF-Transformer, and LSTM models were used to predict for high frequency, medium frequency and low-frequency components to achieve a better balance between global exploration and local exploitation and improve the convergence and robustness of the prediction process. Finally, the data from electric vehicle charging stations in Region H were employed for empirical analysis. The results show that the proposed model outperforms the benchmark models across multiple evaluation metrics, whose R² can reach 0.994 1. The proposed model is suitable for ultra-short-term EV charging load forecasting and can provide effective support for real-time EV load dispatching and minute-level response applications in charging stations.