Parameter Regionalization Of Uninformed Regional Flood Forecasting Model Based On Machine Learning

At present,the global climate fluctuation is abnormal,and heavy rainfall,floods,and other natural disasters occur frequently,which make people’s property suffer great losses and personal safety also faces threats.The establishment of accurate and perfect flood forecasting and warning system can be of great significance for disaster prevention and mitigation.However,due to various factors,the forecast and early warning system in China is not yet perfect,and flood forecasting has not yet been carried out in a considerable number of regions.Especially in the small and medium-sized watershed areas where there is a shortage of information,the accuracy of flood forecasting is facing a great challenge.In this paper,small and medium-sized basins in the Xiangjiang River basin of Hunan Province are selected for analysis,and their production and confluence mechanisms and parameter laws are studied first.Subsequently,the Xinjiang River model is used to parameterize the flood fields of selected hydrological stations and carry out mutual transplantation,and the random forest method(RF),BP neuron method(BP),support vector regression(SVR),and decision tree(CART)based on machine learning are used to quantify the parameter transplantation indexes to find a suitable parameter region transplantation scheme and compare the transplantation effect with that of traditional methods such as similar distance and similar attributes.The purpose of this study is to study the characteristics of flow production mechanism in small and medium-sized basins of Xiangjiang River basin in Hunan Province using machine learning,and offer the theoretical foundation and provide technological support for the regionalization of parameters in areas without data.The research content of this paper is as follows:(1)In this study,20 small and medium-sized watersheds in the Xiangjiang River basin of Hunan Province were taken as the main subjects,and a distributed hydrological model was constructed,and the Xin’an River model was used to simulate the production and confluence,and the Muskingum method was used for river confluence.Long-series hydrological observations were collected to rate and validate the constructed hydrological model.Sobol sensitivity analysis was performed to determine the sensitivity of the model parameters,followed by transplantation tests of the yield flow parameters between watersheds.(2)Due to the large differences in different parametric regionalization methods applied in different watersheds,this study used the attribute characteristics of each watershed to construct parameter regionalization schemes using the random forest method(RF),BP neuron method,support vector regression(SVR),and decision tree(CART)and compared these four schemes with the transplantation results obtained by the similar distance and attribute method.The study showed that the average Nash coefficients of six regionalization schemes using the random forest method(RF),BP neuron method,support vector regression(SVR),decision tree(CART),distance proximity,and attribute similarity method were 0.722,0.717,0.707,0.714,0.691,and 0.580,respectively.The Nash coefficients of the parameter shift values of the distance proximity and attribute similarity schemes are relatively low and the mean square deviations are high,0.009 and 0.041,respectively,which are much higher than 0.001 for the random forest method(RF),BP neuron method,support vector regression(SVR),and decision tree(CART).This indicates that the within-group simulations in the distance proximity and attribute similarity schemes vary widely and have relatively poor overall results,while the machine learning-based random forest method(RF),BP neuron method,support vector regression(SVR),and decision tree(CART)have better overall results and perform more consistently.