基于指数核函数高斯过程回归的短期径流预测研究  

Short-term Runoff Prediction Based on Exponential Kernel Gaussian Process Regression

作　　者：何中政 方丽[1] 刘万 迟英凡 王永强[2] 付吉斯 熊斌[1,3] HE Zhong-zheng;FANG Li;LIU Wan;CHI Ying-fan;WANG Yong-qiang;FU Ji-si;XIONG Bin(School of Architectural Engineering,Nanchang University,Nanchang 330031,Jiangxi Province,China;Changjiang River Scientific Research Institute of Changjiang Water Resources Commission,Wuhan 430010,Hubei Province,China;Key Laboratory of Poyang Lake)

机构地区：[1]南昌大学工程建设学院,江西南昌330031 [2]长江水利委员会长江科学院,湖北武汉430010 [3]南昌大学鄱阳湖环境与资源利用教育部重点实验室,江西南昌330031

出　　处：《中国农村水利水电》2023年第8期25-31,40,共8页China Rural Water and Hydropower

基　　金：国家自然科学基金(52209024);中央基本科研业务费(CKSF2021486/SZ);江西省水利厅科技项目(202223YBKT43);江西省自然科学基金(20212BAB214065);江西省水利科学院开放研究基金(2021SKSH01,2021SKSH02)。

摘　　要：径流预测有助于流域水资源综合高效调配和防洪减灾调度,如何精准地开展短期径流预测一直是水文水资源研究领域的重点。高斯过程回归(GPR)凭借其针对复杂非线性回归问题的泛化能力,已在水文过程长中短期预测研究中得到成功应用。而GPR回归分析能力不仅取决于模型参数,还受核函数影响。为此,研究分析了不同核函数作用下GPR预测模型效果,提出了基于指数核函数GPR的流域短期径流预测模型。首先通过多重相关性系数分析筛选相关性系数大且时段最短的预测因子组合,然后分别选用有理二次、径向基、马顿和指数核函数建立不同的GPR短期径流预测模型,同时加入了MLR、RT、SVM、BP等模型方法的预测结果作为对比。以赣江流域吉安水文站短期径流预测(预测步长为6 h,预见期为7 d)为例开展实例分析,相关实验结果表明:①应用不同核函数的GPR模型预测结果表现存在明显差异,不同方法预测表现由好到差分别为指数GPR、有理二次GPR、RT、马顿GPR、径向基GPR、SVM、MLR、BP;②指数GPR预测模型28时段的4项评价指标均表现最佳,DC和QR分别接近1和100%,预报精度达到甲级以上。综上,研究验证了指数核函数GPR短期径流预测模型的有效性和普适性,模型预测精度满足实际工程应用需求,具备实际应用价值。Runoff prediction is helpful to the comprehensive and efficient allocation of water resources and flood control and disaster reduction operation in the basin.How to accurately carry out a short-term runoff prediction has always been the focus of hydrology and water resource research.Gaussian process regression(GPR)has been successfully applied in the long,medium and short-term hydrological process prediction research because of its generalization ability for complex nonlinear regression problems.The GPR regression analysis ability depends on not only the model parameters but also the kernel function.Therefore,this paper analyzes the effect of GPR prediction model under different kernel functions,and proposes a short-term runoff prediction model based on exponential kernel function.Through the multiple correlation coefficient analysis of the largest multiple correlation coefficient,and the shortest predictor period,and then the rational secondary,radial base,maton and exponential kernel function are chosen to establish different GPR short-term runoff prediction model,also joined the MLR,RT,SVM,BP model method prediction results as a comparison.Taking the short-term runoff prediction of Ji′an Hydrologic Station in the Ganjiang River Basin(the prediction step is 6 h,and the prediction period is 7 days)as an example,the relevant experimental results show that:①There are obvious differences in the prediction results of GPR models using different kernel functions,and the prediction performance of different methods from good to bad is exponential GPR,rational quadratic GPR,RT,Marton GPR,Radial basis GPR,SVM,MLR,BP;②The 4 evaluation indexes of the exponential GPR prediction model in 28 periods all performed best,DC and QR are close to 1 and 100%respectively,the forecast accuracy reaches grade A or above.In conclusion,this paper verifies the effectiveness and universality of the exponential kernel function GPR short-term runoff prediction model,and the model prediction accuracy meets the needs of practical engineeri

关 键 词：短期径流预测 高斯过程回归 指数核函数 赣江流域 吉安水文站 

分 类 号：TV124[水利工程—水文学及水资源]