Study on Forecasting and Alarming Model of Flash Flood Based on Machine Learning

Abstract

AbstractIn the past decade, more than 300 people die a year due to mountain torrents in china on average. Flood forecast and early warning is an important method of mountain torrents prevention. In this paper, machine learning method is used to improve the problem of low accuracy of flood simulation and early warning in ungauged areas. The distributed model covering Hunan Province is constructed as the case study. The SCE-UA algorithm is used to automatically optimize the model parameters and establish the model parameter database based on the data of 30 hydrological stations. The parameters were transplanted according to the terrain, landform, soil and vegetation characteristics of each basin by using shortest distance, attribute similarity and random forest model (RF), and the warning model was calibrated by using support vector machine (SVM). The results show that the regionalization scheme based on the random forest model can improve the average Nash–Sutcliffe coefficient of flood simulation in ungauged areas by 20%, and the accuracy of warning model based on support vector machine has been improved year by year, it will reach more than 50% after running 5–10 years.KeywordsFlood forecast and early warningDistributed hydrological modelMachine learningRandom forestSupport vector machine