This paper addresses the importance of accurately predicting key system parameters such as channel state information (CSI), user locations, and network traffic in complex and dynamic mobile communication networks. We note that existing deep learning (DL)-based methods struggle to generalize across diverse scenarios and tasks, and thus propose a unified base model for multi-task prediction in wireless networks supporting various prediction intervals. The proposed model enhances univariate decomposition to unify heterogeneous tasks, encodes granularity for interval awareness, and uses a causal Transformer backbone for accurate prediction. In addition, we introduce a patch masking strategy during training to support arbitrary input lengths. After being trained on a large dataset, the proposed base model demonstrates strong generalization to unseen scenarios and achieves better zero-shot performance on novel tasks than existing full-shot baselines.
