|国家科技期刊平台
首页|期刊导航|水利水电技术(中英文)|基于机器学习的雨型分类研究:以淠河流域为例

基于机器学习的雨型分类研究:以淠河流域为例OACSTPCD

Research of rain patterns classification based on machine learning:A case study in Pi River Basin

中文摘要英文摘要

[目的]为了提升洪水预报方案的科学性和精度,开展降雨雨型分类,制定不同雨型的预报方案并实施作业预报是一条有效的技术路线.[方法]基于淠河流域2003-2021年37个雨量站逐小时降雨观测数据,利用业界公认的动态时间规划(DTW)算法进行场次降雨雨型分类并作为基准分类结果.在此基础上,分别选取决策树(DT)、长短期记忆神经网络(LSTM)、LightGBM、支持向量机(SVM)四种机器学习方法建立分类模型并检验分类效果.通过调整样本规模,分析不同样本容量对分类效果的影响.[结果]结果表明:四种分类模型中,LightGBM分类精度最高且训练速度快,LSTM和SVM分类精度良好但训练效率相对较低,DT方法分类速度较快但分类精度相对较低.随着样本规模的增大,分类结果逐步稳定,四种方法的分类效果和训练效率逐步提升.[结论]结果验证了机器学习方法在降雨序列雨型分类中具有较强的适用性,可为洪水预报方案分类构建提供技术支撑.

[Objective]In order to enhance the scientificity and accuracy of flood forecasting scheme,It is an effective technical approach to conduct rainfall pattern classification,formulate different rainfall pattern forecasting schemes,and implement opera-tional forecasting.[Methods]Based on hourly rainfall observation data from 37 rain gauge stations in the Pi River Basin during the period of year 2003-2021,the widely recognized Dynamic Time Warping(DTW)algorithm is employed for rainfall pattern classification,and it serves as the benchmark classification result.Subsequently,four machine learning method,namely decision tree(DT),long short-term memory neural network(LSTM),LightGBM,and support vector machine(SVM),are selected to build classification models and evaluate their classification performances.By adjusting the sample size,the impact of different sample capacities on the classification effectiveness is analyzed.[Results]The result reveal that among the four classification models,LightGBM exhibites the highest accuracy and fastest training speed,while LSTM and SVM demonstrate good classifica-tion accuracy but relatively lower training efficiency,and DT exhibites relatively faster classification speed but lower accuracy.As the sample size increases,the classification result gradually stabilize,and the classification effectiveness and training efficiency of the four method improve gradually.[Conclusion]This result validate the strong applicability of machine learning method in rain-fall sequence pattern classification,providing technical support for the classification construction of flood forecasting schemes.

付晓娣;阚光远;刘荣华;梁珂

中国水利水电科学研究院,北京 100038||水利部防洪抗旱减灾工程技术研究中心,北京 100038||流域水循环模拟与调控国家重点实验室,北京 100038||水利部京津冀水安全保障重点实验室,北京 100038北京中水科工程集团有限公司,北京 100048

水利科学

降雨雨型时空分布特征动态时间规划LightGBMLSTM降雨机器学习

rain patternsspatial and temporal distribution characteristicsdynamic time planningLightGBMLSTMrainfallmachine learning

《水利水电技术(中英文)》 2024 (002)

101-111 / 11

国家自然科学基金(42271095);中国水利水电科学研究院十四五"五大人才"计划(JZ0199A032021);GHFUND A(ghfund202302018283);城市水循环与海绵城市技术北京市重点实验室开放基金(HYD2020OF02)

10.13928/j.cnki.wrahe.2024.02.009

评论