In view of the nonlinear and non-stationary characteristics of runoff processes and low prediction accuracy, a daily runoff prediction model combining time-varying-filter-based empirical mode decomposition (TVF-EMD), grey relation analysis (GRA), and light gradient boosting machine (LightGBM) was proposed. Taking the measured daily runoff series of Lijin station on the Yellow River and Gaoyao station on the Pearl River as examples, the TVF-EMD-GRA-LightGBM (TGL) combined model was established, and its prediction results were compared with those of other single or combined models. The results show that the TGL combined model is efficient and has the best prediction performance, and the prediction results of the TGL combined model for Lijin and Gaoyao stations have the Nash-Sutcliffe efficiency coefficient of 0.95 and 0.97, respectively, the correlation coefficient of 0.97 and 0.98, respectively, and the peak flow prediction error less than 0.078 and 0.073, respectively. The TGL combined model has the advantages of high prediction accuracy, high operation efficiency, and strong applicability, providing a new way for daily runoff prediction and scientific regulation of water resources system.