Abstract: Predicting future network status is a key capability in digital twin networks and can assist operators in estimating network performance and taking proactive measures in advance. Existing methods, including statistical methods, machine learning-based methods, and deep learning-based methods, suffer from several limitations in generalization ability and data dependency. To overcome these drawbacks, inspired by the success of pretraining and the fine-tuning framework in the natural language processing and computer vision domains, in this work, we propose TSNet, a transformer-based foundation model for predicting various network status measurements in digital twins. To adapt transformer architecture to time series data, frequency learning attention and time series decomposition blocks are implemented. We also design a fine-tuning strategy to enable TSNet to adapt to new data or scenarios. Experiments demonstrate that zero-shot prediction using TSNet without any training data performs better than fully supervised baseline methods and achieves higher accuracy. In addition, the prediction accuracy could be further improved by equipping the proposed model with fine-tuning. Overall, TSNet exhibits strong capability and accuracy in various datasets. For future work, we plan to explore TSNetâĂŹs generalization abilities on various downstream tasks (such as anomaly detection and pattern matching) using few-shot learning approaches