Abstract: Aiming at the problem of limited correction accuracy of traditional tropospheric delay models, a tropospheric delay prediction model by optimized time convolutional attention network (TCAN) is proposed. This model takes the station location information, time information, and the time series of tropospheric zenith total delay (ZTD) as input features, and the time series of ZTD as output features. Three target windows of 2 h, 6 h, and 12 h are designed for model training, and a short-period ZTD prediction model is established. Taking the continuous 30-day data of 19 continuously operating reference stations (CORS) stations in Hong Kong, China as the research object, the experimental results show that when a 2-hour target window is set, the accuracy of the short-period ZTD prediction model based on the TCAN is better than that of 6 h and 12 h. Bias is 2.28 cm, root mean square error (RMSE) is 3.10 cm, and standard deviation (STD) is 2.11 cm. Compared with the traditional global pressure and temperature3 (GPT3) model, the prediction accuracy of the optimized TCAN model has increased by 124%, 69%, and 48%, respectively. In conclusion, the proposed model has certain application value in the field of the GNSS precise positioning.