Abstract
This paper proposes a hybrid forecasting approach that integrates LSTM neural networks with classical hydrological models to improve the accuracy of early flood prediction. In the course of the study, the Saint-Venant equations were employed to describe the dynamics of open-channel flow, while the Boussinesq equations were used to model groundwater movement. The results of physical-based modeling, together with time-series data, were utilized to train the LSTM neural network. The performance of the proposed hybrid model was evaluated using the RMSE and MAE metrics and compared with the results of standalone classical and data-driven models. The results demonstrate that the hybrid approach provides higher accuracy and robustness in flood forecasting. This study can serve as a basis for early flood risk detection and its application in intelligent management systems.