The Northeast China Cold Vortex (NCCV) is a crucial local synoptic system influencing the weather and climate of Northeast China. However, the application of artificial intelligence techniques in NCCV prediction remains limited. Based on ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF), this study constructs a 23-year multi-modal spatiotemporal sequence dataset of NCCV via an objective identification method, focusing on NCCV trajectory prediction. An improved generative adversarial network model is proposed, which adopts a multi-encoder architecture to extract spatiotemporal features of multi-modal NCCV data and introduces a multi-generator structure to address the insufficient prediction capability of a single generator. A selector module is added to enable the model to adaptively select the optimal generation path. Ablation experiments show that compared with single-trajectory data input, multi-modal data input in our model reduces the average prediction error by 67.96 km, representing a 34.0% improvement, and the 24-h prediction error improvement reaches 39.7%. Ultimately, the proposed model achieves superior prediction accuracy and stability in the NCCV trajectory prediction tasks at 6 h, 12 h, 18 h, and 24 h, with prediction distance errors reduced by 21.4%, 29.2%, 34.0%, and 37.0% compared to LSTM.