Abstract: To address the forecasting challenges of electric vehicle charging loads characterized by strong nonlinearity, high volatility, and multi-feature coupling, this study proposes a multivariate time-series load forecasting framework integrating data decoupling reconstruction and multi-branch PatchMixer. Firstly, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm combined with permutation entropy and k-means clustering is employed to realize dimensionality reduction and reorganization of the original load series. Secondly, trigonometric encoding is introduced to eliminate boundary distance errors of temporal cycles, and a multidimensional feature set is constructed by incorporating meteorological data and other information. Subsequently, light gradient boosting machine (LightGBM) is utilized to evaluate feature importance and screen out redundant features. Finally, a PatchMixer forecasting model based on pure multi-layer perceptron network is constructed. Parallel deep mixing branches are designed for different frequency bands to efficiently capture variation trends of different frequencies. Experimental results demonstrate that the proposed forecasting framework significantly outperforms mainstream baseline models in multiple forecasting metrics, thereby validating its superiority.