<p>Global navigation satellite system-derived precipitable water vapor is a key ingredient in atmospheric water vapor transport and condensation, serving as a critical indicator of rainfall occurrence and intensity variation. However, traditional methods often rely on fixed thresholds, neglecting the temporal dynamics and evolving patterns of water vapor. To address the limitation, we propose ATFcast, an adaptive time-frequency fusion model for rainfall forecasting using GNSS and meteorological data. The model employs the stacked LSTM as its baseline, augmented with two novel components: a frequency domain transducer and a temporal-frequency feature extraction module. We rigorously evaluate the model’s performance in binary classification and quantitative forecasting tasks. The assessment is conducted across mainland China, a region recognized for its complex climate zones and significant geographical diversity. Experimental results unequivocally demonstrate ATFcast’s superior performance over the baseline. Specifically, the accuracy is improved by 1.79% points, the F1 by 2.75% points, the RMSE is reduced by 3.90%, and the MAE is reduced by 2.83%. Furthermore, in a 6-step iterative forecasting task, the RMSE and the MAE are reduced by 8.46% and 8.98%, respectively. Finally, the ablation experiments and SHAP-based interpretability analyses validate the individual contribution of each model component and confirm that ATFcast provides a reliable and adaptive framework for station-level, short-term rainfall forecasting across mainland China.</p>