基于预报因子聚类分级的日径流预报深度信念模型及应用  

Daily Streamflow Forecasting Using Deep Belief Networks Model Based on Predictors Clustering Classification

作　　者：初海波 魏加华[1,2] 王东方 李家叶[1] 黄跃飞[1,2] 李铁键[1,2] CHU Haibo;WEI Jiahua;WANG Dongfang;LI Jiaye;HUANG Yuefei;LI Tiejian(State Key Laboratory of Hydroscience and Engineering,Tsinghua University,Beijing 100084,China;State Key Laboratory of Plateau Ecology and Agriculture,Qinghai University,Xining 810016,China;State Grid Qinghai Electric Power Company,Xining 810016,China)

机构地区：[1]清华大学水沙科学与水利水电工程国家重点实验室,北京100084 [2]青海大学三江源生态与高原农牧业国家重点实验室,青海西宁810016 [3]国网青海省电力公司,青海西宁810008

出　　处：《应用基础与工程科学学报》2018年第5期929-939,共11页Journal of Basic Science and Engineering

基　　金：十三五国家重点研发计划项目(2017YFC0403600);国家电网公司项目(52283014000T);四川省科技厅科技计划项目(2015JZ00110)

摘　　要：径流预报是水库运行调度的重要决策依据,提高入库径流预报精度,对水库优化调度和水资源高效利用具有重要意义.本文提出一种基于大样本数据分级策略的深度信念网络模型（Deep Belief Networks,DBN）,以龙羊峡入库径流预报为例,采用Fuzzy C-means（FCM）聚类方法,将总样本训练集分为不同训练样本子集;不同样本子集下,对不同预报因子（只考虑降雨、考虑不同时期的降雨及同时考虑降雨及前期径流）,分别建立DBN模型和人工神经网络模型（Artificial Neural Network,ANN）,分析样本分级和考虑不同因子情况下不同模型的预报结果.结果表明：与不考虑预报因子分级的预报模型相比,基于时间序列聚类的预报模型显著提高了径流预报的精度;通过FCM聚类,将样本分为3类,考虑降雨及前期径流作为预报因子进行分级时,比只考虑降雨、考虑不同时期的降雨时建立的预报模型的预测精度更高.用该模型进行龙羊峡水库入流日径流预报,提高了预报精度,可为龙羊峡水库调度提供决策支持.Streamflow forecasting is important for optimal reservoir operation and management,and improving the efficiency of water utilization. In this paper,we propose a predictors classification strategy with large amounts of samples,that is,Deep Belief Networks（ DBN） based on Fuzzy Cmeans（ FCM） clustering method. Daily streamflow forecasting of Longyangxia reservoir was taken as the case study. According to FCM clustering method,the total dataset is divided into different sub-datasets. Considering different forecasting factors（ only rainfall,rainfall in different periods,both rainfall and previous streamflow）,DBN and ANN models were built,and the results were compared to analyze the impact of samples classification and different forecasting factors on the accuracy of the forecasting models. Results showed that the forecasting models based on predictor classification significantly improve the accuracy of streamflow forecasting compared to the traditional forecasting models without samples classification. The total samples were classified into three categories by FCM clustering,and the performance of the models considering the rainfall and previous streamflow as the predictors is better than those of the models considering only the rainfall or rainfall in different periods. Streamflow forecasting in Longyangxia reservoir has been improved the accuracy and provide the decision support for the scheduling of Longyangxia reservoir.

关 键 词：预测因子分级 FCM聚类 深度信念网络 径流预报 龙羊峡水库 

分 类 号：P338[天文地球—水文科学]