Calculation of reserves flood routing plays a significant role in joint flood control of reservoirs and river channels. Application of the Muskingum method for reverse calculation may lead to unstable results and poor accuracy, making it unsuitable for practical engineering applications. This paper proposes a flood inverse-routing method based on the Long Short-Term Memory (LSTM) neural network model. This paper establishes a simulation model of the nonlinear relationship between flow in the upstream and downstream sections of the river channel, and trains the model using the historical flood data, thereby achieving the retrograde estimation of the upstream inflow process from the downstream flood process. The Han River was selected for the case study. Results show that the outcomes of the flood inverse-routing method based on the LSTM neural network model closely align with the actual inflow process at the upstream cross-section and outperform those obtained by the BP neural network and Support Vector Regression inversion methods. This demonstrates the practicality and effectiveness of the proposed model.