动态权值自适应分配的LSTM洪水演进方法OA

To enhance the simulation accuracy of flood routing and improve the model's capability to capture flood peaks and adjacent flow variations,a dynamic weight adaptive allocation-based long short-term memory(LSTM)method for flood routing was proposed.First,a dynamic weight allocation matrix was established based on the regularization of prediction error values for flood sub-sequences.Subsequently,the dynamic weights were incorporated into the mean squared error(MSE)loss function to formulate a model optimization strategy with adaptive learning weights.Such strategy enables the model to adaptively amplify learning weights for high-flow periods while attenuating those for low-flow periods.To validate the proposed method,a case study was carried out at the lower reaches of the Hanjiang River from Huangzhuang to Xiantao hydrologic station.Results demonstrated that,compared to typical LSTM approaches,the proposed method significantly improved simulation accuracy for flood peaks and high-flow periods while exhibiting more stable performance across the entire flood process,markedly enhancing the overall effectiveness of flood routing.