Abstract: The total electron content (TEC) of the ionosphere is a critical parameter in radio wave propagation and space activities, making the development of high-precision TEC prediction models highly significant. This paper proposes a combined prediction model (TCN-STT) that integrates spatio-temporal transformer (STT) and temporal convolutional network (TCN) with a spatio-temporal attention mechanism for TEC prediction. The model is based on TEC data from the International GNSS Service (IGS) European Reference Station for a period spanning from 2000 to 2023, covering a total of 8766 days. A sliding window approach was used to generate 8764 samples. These samples were categorized into three groups based on the Kp geomagnetic index (Kp < 4, 4 ≤ Kp < 7, Kp ≥ 7), and random sampling was employed to ensure that the distribution of samples with varying geomagnetic activity levels was relatively balanced across the training, validation, and test sets. The final dataset was split in an 8:1:1 ratio. Experimental results indicate that during geomagnetic quiet periods (Kp < 4), the average RMSE of the samples is 2.62 TECU, with an average relative accuracy of 90.5%. During geomagnetic active periods (4 ≤ Kp < 7), the RMSE increases to 3.94 TECU, and the average relative accuracy drops to 87.7%. In geomagnetic disturbed periods (Kp ≥ 7), the RMSE further rises to 8.95 TECU, with the average relative accuracy decreasing to 81.3%. Overall, the model achieved an average RMSE of 2.68 TECU and an average relative accuracy of 90.36% across all test set samples. Moreover, the correlation coefficient between the predicted and true values on the entire test set is 0.9866, and the coefficient of determination (R²) is 0.9734, demonstrating that the model delivers excellent and stable predictive performance.