Author(s): Wu Biqiong; Zhao Jie; Zhang Hairong; Cao Hui; Bao Zhengfeng; Zhang Dongjie; Zhou Xiangxi

Linked Author(s):

Keywords: Rainfall-runoff; Similarity; Data mining; Machine learning

Abstract: In traditional flood forecasting, forecasters often rely on experience to find historical rainfall processes that are similar to future rainfall forecasts, such as finding historical rainfall runoff that similar with total amount of precipitation, and using these similar historical processes as the basis for correcting forecast model results or directly forecasting. However, this approach is not only time-consuming but also carries a degree of subjectivity and limitations. Therefore, it is of great significance to model the years of experience of forecasters,construct a scientific and comprehensive rain and flood similarity judgment method for identifying the similarity between historical and future rainfall-runoff, and automatically realize real-time flood forecasting. In this study, data mining and machine learning are used to propose a set of comprehensive similarity evaluation methods based on the hydrological state of the basin, the temporal distribution of rainstorm, and the spatial distribution of rainstorm based on the mechanism of rainfall runoff response.The results showed that the average peak and volume errors of the most similar process found by the comprehensive similarity method were significantly reduced compared to the single-sided similarity method, with the maximum error reduced by nearly 90%. This study successfully models and digitizes expert experience, and through comprehensive similarity, can identify historical rainfall events that are similar to the predicted future rainfall process, which has important reference value for real-time flood process forecasting.

DOI: https://doi.org/10.64697/IAHR-APD2024_P404

Year: 2024