Wireless channels exhibit rapid time-varying characteristics due to multipath effects and Doppler shift. which is particularly evident in dynamic monitoring scenarios involving smart textiles and wearable sensing nodes. Traditional estimation algorithms and prediction models suffer from insufficient accuracy and difficulty in capturing global dependencies, failing to meet the demands of modern wireless communication. Therefore, this study aims to construct a high-precision, low-latency CSI prediction model to improve the accuracy and real-time performance of CSI prediction in rapidly changing scenarios, providing reliable support for adaptive transmission. By collecting CSI data in various environments (urban areas, mountainous areas, and open spaces) and preprocessing it through denoising and standardization to construct time-series samples, an improved Transformer model based on a sparse attention mechanism to optimize the encoder-decoder architecture is designed. The optimal hyperparameters are determined by combining grid search and Bayesian optimization. The model is then compared with benchmark models such as linear regression, LSTM, and CNN, and performance is evaluated using multiple metrics including RMSE, MAE, and correlation coefficient. Experimental results show that the proposed Transformer model has a prediction mean square error (MSE) as low as 0.0185 and a correlation coefficient of 0.942, which improves the prediction accuracy by 15%-25% compared with the traditional model. Furthermore, structural optimization effectively balances computational complexity and realtime requirements. This model solves the CSI prediction problem of fast time-varying channels, providing a new path for the intelligent adaptation of wireless communication systems and has significant practical value for optimizing communication performance in 5G/6G.

smart textiles, channel state information, transformer, wireless communication, prediction model