摘要
水资源短缺和生态环境恶化问题，已成为制约我国西北内陆河流域经济可持续发展的关键因素。而随着社会经济的不断发展，国民经济各部门对水文预报提出的要求越来越高，不仅要求有较高精度的短期预测，而且要求有预见期更长的中长期径流预测。因此基于组合模型的中长期径流预测研究，对流域水资源有效调控和合理利用水资源具有重要意义。论文在应用单项预测方法的基础上，应用径流预测的多种组合预测方法预测石羊河流域支流枯季径流和年径流的预测，主要研究内容和取得的主要结论如下：
（1）归纳总结了目前径流预测的主要方法，包括多元线性回归（MLR）、BP神经网络、支持向量机（SVM）和自回归移动平均（ARIMA）四种单项预测方法。为了提高年径流预测的准确性，介绍了简单平均方法（SA）、加权平均方法（WA）、回归方法（Regression）、误差平方和（SSE）、人工神经网络（ANN）和时变权重方法（包括时变误差平方和（TSSE）、线性时变误差平方和（LTSSE）、几何时变误差平方和（GTSSE））八种组合预测技术。通过组合单个预测模型的误差信息，得出精度更高的预测结果。
（2）采用多元线性回归（MLR）、BP神经网络和自回归移动平均（ARIMA）三个单项预测方法对石羊河支流西营河枯季径流进行了预测，用相对均方根误差（R-RMSE）和相对偏差（R-Bias）检验模型的拟合精度，结果表明单项预测模型中ARIMA模型的精度最高；采用加权平均组合预测的精度要高于简单平均组合预测的精度，并且ARIMA-MLR和ARIMA-BP组合的拟合精度较好。
（3）根据前期的气象资料采用多元线性回归（MLR）、BP神经网络、支持向量机（SVM）和自回归移动平均（ARIMA）四种单项预测方法预测石羊河流域八条山水河流出山口的年径流，用R-RMSE和R-Bias检验模型的拟合精度，结果表明，SVM方法的精度最高。
（4）采用八种组合预测方法，组合年径流的四种单项预测模型，共有11种组合可能，并采用预测结果的相对均方根误差（R-RMSE）和相对偏差（R-Bias）比较模型的精度。在理论上说明了：①当SA和WA两个预测误差方差的比率很大，并且当一个预测相对与另一个较差时，WA要比SA表现得好。尽管SA并不总能比那些最好的单个预测模型的精度高，但却不失为一种较为有效的组合技术。②当单项预测具有非稳定性的误差时，时变权重组合方法比恒定权重的方法产生更好的结果。③当两个预测方法的结果高度相关时，组合技术不能带来显著性的改进。在这种情况下，建议应用最好的单项预测。此时，ANN组合方法比线性组合方法可以更精确的代表单个预测之间复杂的和非线性的关系，它的R-RMSE值相对其它组合方法更小。④基于Regression和ANN的组合方法可以去除组合预测中的偏差的影响，应用偏差校正技术相对其它没有偏差校正的技术可以改进组合预测结果。
（5）通过对各种组合预测方法的相对均方根误差比较，获得了八条支流径流组合预测的最优组合方法。大靖河最优的组合是在ANN组合方法下的M*SA组合；古浪河最优的组合是在LTSSE和GTSSE组合方法下的M*SA组合；黄羊河最优的组合是在ANN组合方法下的MB*SA组合；杂木河最优的组合是在Regression组合方法下的M*BSA组合；金塔河最优的组合是在Regression和ANN组合方法下的MB*SA组合；西营河最优的组合是在TSSE、LTSSE和GTSSE组合方法下的M*SA组合；东大河最优的组合是在WA、Regression和ANN组合方法下的M*BS组合；西大河最优的组合方法是在WA、Regression和ANN组合方法下的M*BSA。总体上来说：ANN、Regression和WA组合方法的的效果最好，当误差序列不平稳时，TSSE、LTSSE和GTSSE组合方法可提高精度；而M*BSA组合、MB*SA组合和SA组合的精度最高。
The problems of shortage in water resources and ecological environment deteriorationhas become a key factor in restricting the sustainable development of northwest inland rivereconomic in China. With the continuous development of the socio-economic, the requirementof various departments of the national economy on the hydrological forecasting is gettinghigher and higher, not only requires short-term forecasts of higher accuracy, but requireslong-term stream-flow forecasting. Therefore, long-term stream-flow forecasting which isbased on the combination model is of great significance to the effective regulation of the riverbasin water resources and rational use of water resources. This paper based on the applicationof individual prediction methods, applied various combination forecasting methods ofstream-flow forecasting to predict the Shiyang River Basin tributaries drought periodstream-flow and the annual stream-flow, The main research contents and the main conclusionsobtained as follows:
(1) Summarized the currently main methods of stream-flow forecasting, including fourindividual prediction methods: multiple linear regression (MLR), the BP neural networks,support vector machine (SVM) and autoregressive moving average (ARIMA). In order toimprove the accuracy of the forecasts of annual stream-flow, introduct eight combinationsforecasting techniques: the simple average method (SA), the weighted average method (WA)regression methods (Regression), the sum of squared errors method (SSE), artificial neuralnetwork (ANN) and time-varying weights method (including the time-varying sum of squarederror method (TSSE), linear time-varying sum of squared error (LTSSE) and geometrictime-varying sum of squared error (GTSSE)). We can obtained more accurate predictionresults through constitute the error information of single forecasting model.
(2) Apply three individual prediction methods which is including Mmultiple linearregression (MLR), BP neural network and autoregressive moving average(ARIMA)forecasted the drought period stream-flow of Xiying River of the Shiyang River,Test the fitting accuracy of the models with the relative root mean square error (R-RMSE)and the relative bias (R-Bias), the results showed that the ARIMA method has the highestaccuracy among the single models; The accuracy of the weighted average combination is higher than the simple average combination, and fitting accuracy of the ARIMA-MLRcombination and ARIMA-BP combination are better.
(3) Apply four individual prediction methods which is including multiple linearregression (MLR), the BP neural network, support vector machine (SVM) and theautoregressive moving average (ARIMA) to forecast annual stream-flow from eight landscaperiver of the Shiyang River Basin according to the pre-meteorological data. Test the fittingaccuracy of the models with R-RMSE and R-Bias, the results showed that the SVM methodhas the highest accuracy.
(4) Apply eight combination methods, combinating four individual prediction models ofannual stream-flow, this reached a total of11possible of combinations, and compare theaccuracy of models in relative root mean square error(R-RMSE) and the relative bias(R-Bias)of the prediction results. In theory:①When the forecasting error variance ratio of SA andWA methods, and when a prediction is relative poorer than another, WA method performs wellthan SA method. Although SA does not always perform better than that of the best constituentforecast, it was always much better than that of the worst constituent forecasts.②When theerror of individual forecasting is non-stability, time-varying weights combination methodcould produce better results than the constant weight method.③One cannot expectcombining technique to yield significant improvement when two constituent forecasts arehighly correlated. In such cases, using the best constituent forecast model is recommended. Atthis point, the ANN combining method, which has a more flexible structure and can moreaccurately represent complex and nonlinear relationships between constituent forecasts thanthe linear combining approaches, its R-RMSE is lower than other combination.④Thecombining methods based on regression and ANN can remove the effects of bias in theconstituent forecasts. By applying a technique with a bias correction component, one canimprove the combined forecast over using other techniques without a bias correctioncomponent.
(5) Through compare the R-RMSE of various combinations, obtained the optimalcombinations of Shiyang River eight rivers stream-flow forecasting. The optimal combinationof Dajing River is M*SA combination in ANN methods; The optimal combination of GulangRiver is M*SA combination in LTSSE and GTSSE methods; The optimal combination ofHuangyang River is MB*SA combination in ANN methods; The optimal combination ofZamu River is M*BSA combination in Regression methods; The optimal combination of JintaRiver is MB*SA combination in Regression and ANN methods; The optimal combination ofXiying River is M*SA combination in TSSE、LTSSE and GTSSE methods; The optimalcombination of Dongda River is MB*S combination in WA、Regression and ANN methods; The optimal combination of Xida River is M*BSA combination in WA、Regression and ANNmethods; On the whole: WA、Regression and ANN combination methods perform best, Whenthe error is non-stability, TSSE、LTSSE and GTSSE combination methods can improve theaccuracy; and the accuracy of M*BSA combination、MB*SAcombination and SA combinationare the highest.
引文
曹晓亚,王景堂,王晓梅.1997.山西月径流模拟模型及其应用.山西水利科技,4:9-15
陈福增,邓志平.1989.松花江哈尔滨水文站年径流序列分析与预报,2:6-59
程根伟.1988.对称变换随机月径流模型.四川大学学报(工程科学版),3:58-61
崔景山,叶柳.2009.基于逐步回归均生函数模型的哈尔滨站年径流预测分析.黑龙江水利科技,37:7-9
丁晶,邓育仁,安雪松.1997.人工神经前馈(BP)网络模型用作过渡期径流预测的探索.水电站设计,13:69-74
段召辉,李承军.2004.日径流的组合预测模型.水利水运工程学报,3:67-69
傅新忠,冯利华,陈闻晨.2009.ARIMA与ANN组合预测模型在中长期径流预报中的应用.水资源与水工程学报,20:105-109
顾海燕.2008.河川径流量预测研究.[硕士学位论文].黑龙江哈尔滨:东北林业大学
何新林,宋玲,郑旭荣.1997.径流中长期预报的灰色预测方法.石河子大学学报(自然科学版),1:234-237
胡宾,崔广柏,朱灵芝.2007.BP神经网络预测河流月径流量.浙江水利科技,2:15-16
黄领梅.2005.水文要素对人类活动响应的研究—以和田河流域为例.[硕士学位论文].西安:西安理工大学
黄伟军,赵永龙,丁晶.1996.径流的最优组合预测及其贝叶斯分析.四川大学学报(工程科学版),94:97-102
黄志强,李柏宏,凌仙华.2009.水文模型组合预报应用研究.浙江水利水电专科学校学报,21(1):30-33
季小兵,王志坚,魏光辉.2009.加权马尔可夫链在开都河径流预测中的应用.人民黄河,31:41-44
蒋传文,权先璋,陈实等.1999.径流序列的混沌神经网络预测方法.水电能源科学,17:39-41
金菊良,杨晓华,丁晶.1999.基于神经网络的年径流预测模型.人民长江,30:58-59
李计,李毅,严宝文.2010.基于四种方法的月径流预测研究.水利与建筑工程学报,8:56-59
李杰友,夏红梅,熊学农,林镜榆.2001.投影追踪回归法在月径流长期预报中的应用.河海大学学报,29:46-49
李杰友,熊学农.1998.水库月径流中长期预报系统的开发研究.水利水电科技进展,18:32-34
李俊伟.2008.加权马尔可夫链模型在韩江流域年径流状态定性预测中的应用.广东水利水电,9:23-26
李松仕,吴燕.1989.ARMAX模型及其在径流短期预测中的应用.水利科技,4:17-20
李秀云,汤奇成.1988.西北干旱地区枯水径流分析和预测.冰川冻土,1:23-27
李亚娇,沈冰,李家科.2007.年径流预测的小波系数加权和模型.应用科学学报,25:96-99
李彦彬,尤凤,黄强.2010.多元变量径流预测的最小二乘支持向量机模型.水力发电学报,29:28-33.
刘冀,王本德,周惠成,刁艳芳.2008.月径流的分期组合预报模型及应用.水电能源科学,26:1-4
刘恩宝,张家桢.1985.用逐步回归周期分析预报河川径流年径流量.人民长江,1:20-24
刘金平,张建云.2005.中国水文预报技术的发展与展望.水文,25(6):1-5
刘敏.2007.和田绿洲地下水时空分布规律及其生态环境效应研究.[硕士学位论文].西安:西安理工大学
刘权授,方乐润.1988.多阶自回归月径流随机模型及其应用.水文,5:33-37
刘晓安,王金文,王海伟.2006.基于小波分析的月径流ARIMA预测方法.水电自动化与大坝监测,30:77-80
刘玉涛,杜兆国,姜红.1999.非平稳时间序列随机模型在年径流量预测中的应用.黑龙江水专学报,26:45-47
刘治理,马光文,严秉忠.2005.最近邻一径向基函数网络模型在日径流预测中的应用.水资源研究,26:24-26
邵年华.2010.水文时间序列几种预测方法比较研究.[硕士学位论文].陕西西安:西安理工大学
史银军.2011.基于水资源转化模拟的内陆河流域水资源优化配置研究.[硕士学位论文].陕西杨凌:西北农林科技大学
粟晓玲.2007.石羊河流域面向生态的水资源合理配置理论与模型研究.[博士学位论文].杨凌:西北农林科技大学
孙汉贤,王锐琛.1985.黄河上游年、月径流随机模拟及其应用简介.西北水电,2:23-27
孙天青.2010.秃尾河流域径流变化规律及预测研究.[硕士学位论文].陕西:西北农林科技大学
汤成友,官学文,张世明.2008.现代中长期水文预报方法及其应用.北京:中国水利水电出版社
王博,马跃先,贺北方.1999.月径流序列的多层递阶预报研究.系统工程理论与实践,7:132-135
王朝阳.1989.黑龙江漠河站月径流的随机模拟.东北水利水电,7:8-13
王林旺,郎旭东,刘学文.1995.周期性自回归模型月径流系列的生成.山西水利科技,1:66-71
王守荣,郑水红,程磊.2003.气候变化对西北水循环和水资源影响的研究.气候与环境研究,8(1):43～45
王星星.2009.灌区年径流量预测与灌区优化配水研究.[硕士学位论文].陕西西安:西安建筑科技大学
吴泽宁,丁大发,王振华.1996.流域站年、月径流系列随机生成模型及应用.郑州工业大学学报,17:7-10
许东,吴铮.2002.基于MATLAB6.x的系统分析与设计一神经网络(第二版).西安:西安电子科技大学出版社:20-191
杨荣新.2006.基于组合智能优化算法的云南省径流预报系统研究.[博士学位论文].武汉:华中科技大学
殷峻暹,蒋云钟,鲁帆.2008.基于组合预测模型的水库径流长期预报研究.人民黄河,30:28-29
游海林.2010.大伙房水库径流中长期预报方法应用研究.[硕士学位论文].辽宁:辽宁师范大学
于国荣,叶辉,夏自强,赵小勇.2009.投影寻踪自回归模型在长江径流量预测中的应用.河海大学学报(自然科学版),37:263-266
曾文碧,林立金.2010.基于GM(1,1)模型的青衣江上游月径流预测研究.陕西农业科学,5:92-94
张勃,王海青,张华.2009.基于人工神经网络的莺落峡月径流模拟预测.自然资源学报,24:2169-2177
张江辉,邱胜彬,徐建新.1994.叶尔羌河年径流的灰色预测.新疆水利,6:42-46
张捷斌.1988.时间序列分析在乌鲁木齐河年径流预测中的应用.干旱区地理,4:13-18
张伟.2008.基于人工神经网络的径流预测研究.[硕士学位论文].新疆:石河子大学
张学工.2000.关于统计学习理论与支持向量机.自动化学报,26(l):32-2
赵长森.2005.和田绿洲散耗型水文模型研究与应用.[硕士学位论文].西安:西安理工大学
赵丽娜,宋松柏.2010.BP神经网络在大理河年径流预测中的应用.人民黄河,32:30-31
郑昱.1999.基于小波变换的年径流序列的模拟.黄河水利职业技术学院学报,11:1-5
郑世武,拾兵.2010.基于改进灰色系统理论的年径流量预测研究.人民黄河,32:40-41
邹文安,刘立博,王凤.2008.人工神经网络BP模型在枯季径流量预测中的应用.水资源研究,29:43-45
Bates J M, Granger C W J.1969. The combination of forecasts. Oper. Res. Quart,20:451-468
Butts M B, Payne J T, Kristensen M, Madsen H.2004. An evaluation of the impact of model structure onhydrological modelling uncertainty for streamflow simulation. Journal of Hydrology,298(1-4):242-266.
Charles M B, Douglas R L, Dyan C H.1985. Comparison of two daily streamflow simulation models of analpine watersehed. Journal of Hydrology,77(1-4):171-186
Chen Y H, Chang F J.2009. Evolutionary artificial neural networks for hydrological systems forecasting.Journal of Hydrology,367(1-2):125-137
Clemen R T, Winkler R L.1986. Combining economic forecasts. Journal of Business Economic Statistics,4:39-46
David A P, Anthony J J.1999. Predicting the daily streamflow of ungauged catchments in S.E. Australia byregionalising the parameters of a lumped conceptual rainfall-runoff model. Ecological Modelling,123(2-3):91-104
Diebold F X, Pauly P.1987. Structural change and the combination of forecasts. Journal of Forecast,6:21-40
Donaldson R G, Kamstra M.1996. Forecast combining with neural networks. International Journal ofForecast.15:49-61
Duckstein L.1987. Bayesian forecasting of hydrologic variable under Changing Climatology. IAHSPublications,168:301-304
Figlewski S, Urich T.1983. Optimal aggregation of money supply forecasts: accuracy, profitability andmarket efficiency. Journal of Finance,28:695-710
Goodwin P.2000.Correct or combine? Mechanically integrating judgmental forecasts with statisticalmethods. International Journal of Forecast,16:261-275
Gorantiwer S D, Majumdar M.1988. Application of autoregressive model of first order to annual flow ofsome West Bengal streams. Journal of the Institution of Engineers(India):Civil Engineering Division,69:156-159.
Granger C W J, Newbold P.1977. Forecasting Economic Time Series. New York, US: Academic Press
Granger C W J, Ramanathan R.1984. Improved methods of combining forecasts. Journal of Forecast,3:197-204
Green N M D.1973. A synthetic model for daily streamflow. Journal of Hydrology,20(4):351-364
Guo J, Zhou J Z, Qin H, Zou Qiang, Li Q Q.2011. Monthly streamflow forecasting based on improvedsupport vector machine model. Expert Systems with Applications,38(10):13073-13081.
Haykin S.1994. Neural Networks: A Comprehensive Foundation. New
Jeong D I, Kim Y O.2009. Combining single-value streamflow forecasts–A review and guidelines forselecting techniques. Journal of Hydrology,377:284-299
Jeong D I, Kim Y O.2009. Combining single-value streamflow forecasts–A review and guidelines forselecting techniques. Journal of Hydrology,377:284-299
Kabir R, William W H, Alex J C.2012. Daily streamflow forecasting by machine learning methods withweather and climate inputs. Journal of Hydrology,414-415(11):284-293
Kelman J.1977. Stochastic modeling of Hydrologic, intermittent daily processes. Hydrology Papers,89:234-258
Kim Y O, Jeong D I, Ko I H.2006. Combining rainfall–runoff models for improving ensemblestreamflow prediction. Journal of Hydrology,11(6):578–588
Liang G C.1988. Identification of a multiple input, single output, linear, time invariant model forhydrological forecasting. Journal of Hydrology,101(1-4):251-262
Liu F, Zhou J Z, Fang R C.2005. An improved particle swarm optimization and its application inlong-term streamflow forecast. International Conference on Machine Learning andCybernetics(ICMLC):2913-2918
Maissis A H.1997. Optimal filtering techniques for hydrological fore-casting. Journal of Hydrology,33(3-4):319-330
McLeod A I, Noakes D J, Hipel K W, Thompstone R M.1987. Combining hydrologic forecast. Waterresources planning and management.113(1):29-41
Ni Q W, Wang L, Ye R Z.2010. Evolutionary modeling for streamflow forecasting with minimaldatasets:A case study in the West Malian River, China. Envrionmental Engineering Science,27:377-385
Ozgur K.2010. Wavelet regression model for short-term streamflow forecasting. Journal of Hydrology,389(3-4):344-353
Pike J G.1964. The estimation of annual run-off from meteorological data in a tropical climate. Journal ofHydrology,2(2):116-123
Refsgaard J C, Hansen E.1976. Economic value of low flow data. Nordic Hydrology,7:57-72
Rogelio C L, John W L, Jose D S.1982. Optimal management of multireservoir systems using streamflowforecasts. Developments in Water Science,17:553-580
Scholkopf B, Smola A J, Willianson R et al.2000. New support vector algorithms[J]. Neural Computation,12:1207-1245
Schreider S Y, Jakeman A J, Dyer B G, Francis R I.1997. A combined deterministic and self-adaptivestochastic algorithm for streamflow forecasting with application to catchments of the Upper MurrayBasin, Australia. Environmental Modelling&Software,12(1):93-104
Schweppe F C.1973. Uncertain Dynamic Systems. Prentice-Hall.
Shamseldin A Y, Oconnor K M, Liang G C.1997. Methods for combining the
Sivakumar B.2003. Forecasting monthly streamflow dynamics in the western United States: a nonlineardynamical approach. Environmental Modelling&Software,18(8-9):721-728
Theil H.1971. Applied Economic Forecasting, Amsterdam, North-Holland
Theodore K A, Konstantine P G.1997. Parallel computation for streamflow prediction with distributedhydrologic models. Journal of Hydrology,197(1-4):1-24
Vapnik V, Levin E, Le C Y.1994. Measuring the VC-dimension of a leaning machine. Neural Computotion,6:851一876
Wen W, Pieter H A J M, Van G, Vrijling J K, Jun M.2006. Forecasting daily streamflow using hybridANN models. Journal of Hydrology,324(1-4):383-399
Winkler R L, Clemen R T.1992. Sensitivity of weights in combining forecasts. Opns. Res,40(3),609-614
Winkler R L, Makridakis S.1983. The combination of forecasts. Journal of the Rpyal Statistical SocietySeries A-Statistics in Society.146:150-157
Wong H, Ip W C, Zhang R Q, Xia J.2007. Non-parametric time series models for hydrological forecasting.Journal of Hydrology,332(3-4):337-347York US: MacMillan
Zealand C M, Burn D H, Simonovic S P.1999. Short term streamflow forecasting using artificial neuralnetworks. Journal of Hydrology,214(1-4):32-48