The invention relates to a neural network method for predicting river flow rate based on multi-section water levels. The neural network method includes the steps of acquiring known flow rate data and water level data of a to-be-solved hydrometric station last year as well as water level data of m upstream gauging stations and n downstream gauging stations of the hydrometric station, and numbering the water level data so as to obtain water level data of the gauging stations; enabling the water level data measured at the same time last year, the water level data of the adjacent upstream gauging stations and the water level data of the adjacent downstream gauging stations to form groups of independent variables Hi; subjecting each group of independent variables Hi and flow rate data Qi at the corresponding time to BP neural network fitting; according to a time corresponding to known to-be-solved flow rate data Qm, taking the water level data measured at the time as independent variables Hm; substituting the independent variables Hm into a BP neural network model to obtain the corresponding flow rate data Qm so as to obtain a water level-flow rate relation curve and achieve prediction on the river flow rate.