Abstract: Evolution of river flow is vital for water regulation in the Yellow River basin. There is an urgent need for a high-precision, low-latency simulation model that considers an artificial lateral water cycle to meet the demand for accurate and refined water regulation from water sources to end-users. Based on the analysis of unbalanced water volume in the Yellow River's main stream, this study used the CNN-LSTM algorithm to build flow evolution models. The SCE-UA algorithm was used to tune the parameters. The “store-wet-release-dry” mechanism of the Longyangxia Reservoir was embedded in the globally-calibrated CNN-LSTM model for simulation. Model comparison indicated the following: ① Unbalanced water volume in the main stream is in the order of middle ＞ lower ＞ upper reaches, with no significant sub-interval correlation. ② The comprehensive evaluation indicators (E_R-R-M) mean values of the hydrological method, CNN-LSTM (local optimization), CNN-LSTM (global optimization), and the coupled simulation were 0.627, 0.613, 0.774, and 0.811 respectively. The upper and lower reaches had higher accuracy than the middle reaches. ③ Building a deep neural network guided by physical simulation can effectively limit the solution space, with 29.3% higher accuracy than the hydrological method, which has practical value for water regulation in the Yellow River.