To improve the prediction accuracy of monthly runoff in the upper reaches of the Yangtze River under the background of inconsistent hydrological elements, based on the monthly runoff data from 1961 to 2022 at Yichang and Cuntan hydrological stations, a hybrid prediction model combining complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN), fuzzy entropy (FE), spectral clustering (SC), optimal variational mode decomposition (OVMD), and Transformer deep learning model (CEEMDANFESCOVMDTransformer) was constructed to simulate and predict the monthly runoff in the upper reaches of the Yangtze River.The results showed that the hybrid model performed well in predicting the monthly runoff at Yichang and Cuntan hydrological stations, with NSE up to 0.9 in the training period, and NSE of 0.84 and 0.89 in the testing period, respectively; the CEEMDANFESCOVMD decomposition framework can improve the accuracy of peak flow prediction during the flood season；the twostage decomposition structure of the OVMD can effectively reduce the complexity of the highfrequency sequence of monthly runoff, and improve the stability of runoff prediction.