The middle and upper reaches of Puyang river basin is a tidal river, which is easily affected by the downstream flood, resulting in the unstable relationship between water level and discharge, and that brings difficulty to the flood forecasting operation. This paper takes the basin above Zhuji Station as the research object, and 20 representative floods from 1990 to 2021 are selected to forecast the discharge of Zhuji Station by using the three-water source Xin 'anjiang model. Based on the excellent nonlinear mapping capability of BP neural network, a water level prediction model of Zhuji Station was established, which took Meichi water level, temporal rainfall, the discharge of Anhua reservoir and the forecast discharge of Zhuji Station as input factors, to explore the applicability of this model in Zhuji flood peak water level forecasting. The results show that the forecast effect of the flood peak water level in 2019-2021 is good, and the relative error of water level is kept within 2.5%。