A method is being developed for determining the parameters of hydraulic resistance to water flow in a riverbed by matching the results of direct hydrodynamic modeling with measurement data at gauging stations using machine learning methods. The numerical hydrodynamic model is based on shallow water equations using Combined Smoothed Particle Hydrodynamics—Total Variation Diminishing on the GPU. The hydraulic resistance model contains three free parameters, calculated using a neural network with a Long Short-Term Memory architecture. An additional free parameter determines the boundary conditions at the exit of the river flow from the computational domain. Determination of these four characteristics provides the best match between time series of water levels measured at several gauging stations and numerical modeling data using the example of the Lower Volga River. Two methods for determining characteristics were investigated. In the first case, the parameters maintained their constant values throughout the year. The second approach is based on calculating the non-stationary characteristics of hydraulic resistance for each day.