To improve the accuracy of watershed runoff simulation and prediction under changing environment, a TVGM (time variant gain model)-LSTM (long short-term memory) coupling model was constructed for the Baihe River Basin, and the random forest algorithm was used to identify the optimal explanatory variables of the model. The model was applied to simulate the runoff of the Baihe River Basin from 2011 to 2018, and the results show that the TVGM-LSTM coupling model has a good runoff simulation effect in the Baihe River Basin, with the Nash efficiency coefficient reaching 0.95 and 0.90 in the calibration and validation periods, respectively. Compared with TVGM, the TVGM-LSTM coupling model improves the simulation accuracy of non-flood season runoff, and it performs well in simulating flood peaks in flood and non-flood periods. In addition, the TVGM-LSTM coupling model can effectively avoid the overfitting problem, and it has better generalization performance and stable runoff prediction accuracy.