两类集成学习算法在中长期径流预报中的应用  

Application of Two Kinds of Ensemble Learning Algorithms in Medium and Long Term Runoff Forecast

作　　者：许斌 杨凤根[1] 郦于杰 XU Bin;YANG Fenggen;LI Yujie(School of Earth Sciences and Engineering,Hohai University,Nanjing 211100,Jiangsu,China;College of Hydrology and Water Resources,Hohai University,Nanjing 210098,Jiangsu,China)

机构地区：[1]河海大学地球科学与工程学院,江苏南京211100 [2]河海大学水文水资源学院,江苏南京210098

出　　处：《水力发电》2020年第4期21-24,34,共5页Water Power

基　　金：水利部公益性行业科研专项(201301044)。

摘　　要：将前期130项遥相关气候指数作为预报因子,利用分类回归树算法(CART)作为基学习器,引入基于Bagging算法的随机森林模型(RF)与基于Boosting算法的梯度提升树模型(GBDT)的两类集成学习算法作为强学习预报模型,实现对丹江口水库未来1个月、1季度及1年3类径流序列的滚动预报,并通过相对误差绝对值的平均值(MAPE)、Nash效率系数(NSE)、相对均方根误差(RRMSE)、合格率(QR)等指标进行对比分析。研究结果表明,两类模型在验证期模拟精度相似,结果相仿,误差分布较均匀,可进一步用于集合径流预报;随着预报对象量级的增加,径流序列的不稳定性与极值序列分布的不均匀性得以降低,预报的准确度、可靠度以及稳定度得到提高。Taking 130 remote-related climate indexes as forecast factors,the classification and regression tree(CART)is used as a basic learner,and two kind of ensemble learning algorithms of the random forest(RF)based on Bagging model and the gradient boosting decision tree(GBDT)based on Boosting model are introduced as strong learning forecast models to realize the rolling runoff forecast of Danjiangkou Reservoir in the next month,the next quarter and the next year respectively.The forecast results are comparatively analyzed through the indicators of mean absolute percentage error(MAPE),Nash-Sutcliffe efficiency coefficient(NSE),relative root mean square error(RRMSE)and qualification rate(QR).The results show that,(a)the simulation accuracy of two kinds of models is close,the results are similar and the error distribution is relatively uniform during the verification period,which can be further used for ensemble runoff forecasting;and(b)with the increase of the magnitude of forecast objects,the instability series and the uneven distribution of extreme value series of runoff are reduced,and the accuracy,reliability and stability of runoff forecast are improved.

关 键 词：中长期径流预报 机器学习 集成学习 分类回归树 随机森林 梯度提升树 

分 类 号：TV121.4[水利工程—水文学及水资源]