To improve the accuracy of monthly runoff prediction and to address the problem that the conventional decomposition integrated runoff prediction models incorrectly uses future data, a gated recurrent unit (GRU) monthly runoff prediction model (ASWPD-BO-GRU) based on self-adaptation strategy wavelet packet decomposition (ASWPD) and Bayesian optimization (BO) is proposed and developed. First, in order to reduce the prediction difficulty the original monthly runoff time series is decomposed using ASWPD, by which four relatively regular decomposed subseries are obtained without using future data. Then, the hyperparameters of the GRU model corresponding to the decomposed subseries are optimized using BO. Finally, the monthly runoff prediction results are obtained by predicting each subseries and summing and reorganizing the prediction results. The proposed and established model is applied to the prediction of monthly runoff at Yingluoxia Hydrological Station in the Heihe River Basin, and the prediction results are compared with those of GRU, BO-GRU, and WPD-BO-GRU models (models based on the conventional decomposition idea which decomposes the original monthly runoff time series as a whole). The results show that the Nash-Sutcliffe efficiency coefficient (NSE) of ASWPD-BO-GRU model is 0.89, which has the highest prediction accuracy in the example application, indicating that the ASWPD-BO-GRU model has higher prediction accuracy and stronger generalization ability with correct decomposition.