Wireless communication is in strong demand for high throughput and high reliability, while dynamic scenarios bring challenges to these demands. Prediction of network states estimates the future state of the varying channel, which could provide the necessary information for the receiving terminal to adjust transmission strategies accordingly. We propose a transformer-based model called ULTRA to predict the network states. In the proposed model, self-attention is exploited to pursue features between variates, and a trend extractor is introduced to pursue local and global temporal features. We implement comprehensive experiments and an ablation study. The results with real-world data demonstrate that our model outperforms state-of-the-art models.