Since the existing methods of water level forecasting in the downstream of a hydropower station have a larger computational error, Gezhouba Hydropower Station was taken as a research object to establish a novel method for downstream water level variation forecasting during the non-abandoning water period. Based on the BP neural network and the hydropower station monitoring data, the downstream water level forecasting with high accuracy was realized, which can satisfy the needs of real-time scheduling. Compared with the methods of water level-discharge relationship and the empirical formula for unsteady flows, the present forecasting method has the following advantages: (1)It does not use the storage outflow data to calculate the water level, and the influence of the outflow calculation error can be avoided; (2)The hysteretic nature effects of the downstream water level variation is considered during the model construction process, so the forecast accuracy during the hump modulation periods can be greatly improved; (3)It can forecast the downstream water level changing process directly with stable calculation results and higher calculation accuracy, which can significantly improve the forecast accuracy under large peak shaving operating conditions of Gezhouba Hydropower Station during the non-abandoning water period.