A New Framework Based on Data-Based Mechanistic Model and Forgetting Mechanism for Flood Forecast

The classification and identification can increase the prediction accuracy effectively due to the complexity and regularity of flood formation. However, it is difficult to extract the influence indicators, especially in data-sparse basins. This research proposes a framework for flood classification and dynamic flood forecast identification in data-sparse basins. The framework starts from a new perspective for flood classification and introduces the concept of forgetting mechanism for flood identification. In the framework, the Data-Based Mechanistic (DBM) forecasting model, a data-driven model with a physically mechanistic interpretation, has been selected as the basic simulated model; then a flood classification model based on DBM and the process of flood occurrence and development has been built to classify floods and generate the corresponding sub-cluster models, and the similarity of the process of flood occurrence and development for each flood is described as the similarity of the simulated model trained for each flood; the forgetting mechanism, which can eliminate the out-of-date data gradually to reduce the influence of the misleading information, is coupled with the deterministic coefficient to identify one of the sub-models for the dynamic flood forecast. The framework has been tested in Shihuiyao Basin, Northeastern China. Results show that the average deterministic coefficients of the proposed framework are 0.87 and 0.86, which are 0.05 and 0.16 higher than those without classification and identification (0.82 and 0.70). The established framework provides a new idea for flood classification and identification, which has the advantages of ease of use, good generality, and low data requirements.