An improved three-layer machine learning searching algorithm for regional frequency analysis of extreme rainfall events

According to the physical mechanism of extreme rainfall formation, combined with machine learning techniques such as isometric feature mapping and influence region, an improved algorithm is proposed based on the regional frequency analysis method of three-layer searching method, which reduces the uncertainty of the quantile value of the rainfall intensity-duration-frequency (IDF) model. According to the climate and terrain characteristics of the target site, the existing three-layer searching method uses the feature selection method to select the most representative geographical or meteorological elements as the similarity factors to form a homogeneous population, and then obtain the rainfall intensity quantile value with low uncertainty. Considering the nonlinear correlation between feature elements, feature extraction method and supervised clustering method are added on the basis of the existing three-layer searching method to reduce the influence of the nonlinear correlation between feature elements on homogeneous group clustering and its dependence on the input site data. The method was tested at rainfall stations in British Columbia, Canada, and the test results show that the method can further reduce the uncertainty of IDF estimation results.