基于随机森林与深度神经网络的水库枯季入库径流中长期预报  

Medium-term and Long-term Forecast of Reservoir Runoff in Dry Season Based on Random Forest and Deep Neural Network

作　　者：肖三明 刘涛 XIAO Sanming;LIU Tao(The Water Conservancy and Hydropower Planning and Design Institute of Shanwei,Shanwei 516600,China;Goldwind Science&Technology Company Limited,Beijing 100176,China)

机构地区：[1]汕尾市水利水电建筑工程勘测设计室,广东汕尾516600 [2]金风科技股份有限公司,北京100176

出　　处：《广东水利水电》2023年第7期54-58,共5页Guangdong Water Resources and Hydropower

摘　　要：准确可靠的枯季中长期入库径流预报对于指导水库枯水期开展水量调度等具有重要意义。本文以公平水库为研究对象,首先利用随机森林模型(RF)对水文气象因子进行筛选,然后基于深度神经网络模型(DNN)构建水库枯季入库径流中长期预报方案。结果表明:DNN模型对公平水库枯季中长期径流的模拟结果较好,率定期Nash系数为0.952,验证期为0.774,模型具有较强的泛化能力;次年3月的模拟精度较其他月份更优,受异常海温指数的影响,验证期次年1月的模拟结果较差;由于RF模型筛选预报因子侧重点的不同,当量级增大时,DNN模型出现了模拟结果较小量级时明显偏小的情况。Accurate and reliable medium-long term inflow runoff forecast in dry season is of great significance for guiding the water regulation of reservoirs in dry season.In this paper,the Gongping reservoir is taken as the research object.First,the hydrological and meteorological factors are selected by using the Random Forest(RF)model,and then the medium-long term forecasting scheme of the reservoir's inflow runoff in dry season is constructed based on the Deep Neural Network(DNN)model.The results show that the DNN model has a good simulation result for the medium and long term runoff in the dry season of the Gongping reservoir,with the Nash coefficient of 0.952 and the validation period of 0.774.The model has a strong generalization ability;The simulation accuracy in March of the next year is better than that in other months,and the simulation results in January of the next year are poor due to the influence of abnormal sea temperature index;Due to the different emphasis of RF model screening prediction factors,When the equivalent level increases,the simulation results of DNN model are obviously smaller than those of small magnitude.

关 键 词：枯季径流 机器学习 中长期径流预报 随机森林 深度神经网络 

分 类 号：P338.2[天文地球—水文科学]