Abstract
The current flood forecasting models heavily rely on historical measured data, which is often insufficient for robust predictions due to practical challenges such as high measurement costs and data scarcity. This study introduces a novel hybrid approach that synergistically combines the outputs of traditional physical-based models with historical data to train Long Short-Term Memory (LSTM) networks. Specifically, the NAM hydrological model and the HD hydraulic model are employed to simulate flood processes. Focusing on the Jinhua basin, a typical plains river area in China, this research evaluates the efficacy of LSTM models trained on measured, mixed, and simulated datasets. The LSTM architecture includes multiple layers, with optimized hyperparameters tailored for flood forecasting. Key performance indicators such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Peak-relative Error (PRE) are employed to assess the predictive accuracy of the models. The findings demonstrate that LSTM models trained on mixed datasets with a simulated-to-measured data ratio of less than 2:1 consistently achieve superior performance, exhibiting significantly lower RMSE and MAE values compared to models trained on mixed data with larger data ratios. This highlights the advantage of integrating measured and simulated data, leveraging the strengths of both data types to enhance model accuracy. Despite its advantages, the approach has limitations, including dependence on the quality of simulated data and potential computational complexity. However, the development of this hybrid model marks a significant advancement in flood forecasting, offering a promising solution to the challenges of computational efficiency and data scarcity. Potential applications of this approach include real-time flood prediction and risk management in other flood-prone regions, providing a robust framework for integrating diverse data sources to improve forecasting accuracy.
