<p>With the advent of sixth-generation (6G) mobile networks, the growing complexity of traffic patterns and escalating demands pose significant challenges to traditional prediction models, particularly when discrepancies exist between the training and deployment environments and the data distribution is unknown. To address this issue, we propose the Latent Domain Generalization Prediction Framework (LDGPF), an out-of-distribution (OOD) generalization approach that can be applied to various backbone prediction architectures. Motivated by the heuristic consideration that domain-specific information in time-series data often resides in its high-frequency seasonal components, LDGPF pre-trains a Variational Autoencoder (VAE) on the seasonal component of the time series to enhance its ability to capture domain characteristics. During the unified training phase, the pre-trained domain encoder is used to derive latent domain representations, followed by adversarial learning to extract domain-invariant features. The domain-discriminative features and domain-invariant features are then fused for prediction. At this stage, the complete time series is fed into the domain encoder to ensure that the learned representations are better aligned with the final prediction objective. By learning diverse distributions through latent domain representation learning and subsequently extracting domain-invariant features to reduce domain-specific dependencies, LDGPF effectively mitigates distribution shift. Furthermore, by balancing domain invariance with domain-specific information beneficial to prediction, LDGPF enhances its generalization capability. Experiments conducted on two public mobile traffic datasets demonstrate that LDGPF achieves significant improvements over baseline models in both cross-spatial and cross-temporal tasks, validating its effectiveness and practicality.</p>