Abstract: Most of China's small and medium-sized reservoirs lack measured hydrological data, and simulating inflow is of great significance for the operation, water resource utilization, and management of small and medium-sized reservoirs with scarce hydrological data. Taking rainfall, runoff, evaporation, and underlying surface as characteristic parameters, considering variational mode decomposition (VMD), a simulation model of daily inflow based on improved long short-term memory artificial neural network (LSTM) is established. Two cross strategy training datasets are designed for Baogaidong Station and Meitian Reservoir Station in Liuyang River Basin, Hunan Province, which have long-term measured characteristic parameters. During the training period, the improved LSTM model is cross learned and trained on the two datasets, and then transferred to Fuling Reservoir for inflow runoff simulation, and compared and analyzed with SWAT simulation results. The results showed that: ①The simulation accuracy of the VMD-LSTM model was verified by cross training every 3 years or every 12 years, and the Nash efficiency coefficient of the former reached 0.914, slightly higher than that of the latter; ②The inflow process of Fuling Reservoir based on VMD-LSTM simulation is comparable to the SWAT simulation results, with a correlation coefficient of 0.905. The runoff simulation model based on VMD-LSTM can effectively learn the characteristics of watershed runoff and can provide an effective method for simulating and predicting the inflow of small and medium-sized reservoirs without data.