Mid To Long Term Runoff Forecasting Under Non-stationarity And Its Application In Reservoir Operation

Due to the globally climate change, hydrological series tend to displaynon-stationarity, which brings huge challenges to hydrological predictions. In recent tenyears, the upper and middle reaches of Shule river basin have gone through prominentclimate change, with the runoff continuously increasing. Under non-stationarity,traditional runoff prediction methods cannot be used directly; hydro-climaticteleconnection based runoff prediction may provide a possible way, for its clearphysical meaning usually leads to longer prediction horizon and higher precision.This paper first analyses the characteristics of observed runoff series atChangmabao station, and makes analysis of its recharge sources as well as vaporsources of rainfall. Then it explores the relationship between runoff and global seasurface temperature. Based on the above theoretical analysis, two prediction models arebuilt on the basis of hydro-climatic teleconnection analysis. One is a Relevance VectorMachine model, and the other is a multi-regression model. Precision evaluation of thetwo models turns out to be not perfect. Given the fact that runoff during the droughtperiod is mainly recharged by groundwater and rainfall, the runoff series would displayrelatively strong auto-correlation. Thus, two more models considering runoffauto-correlation are built. One is time series model, the other is a coupling model whichis based on both hydro-climatic teleconnection and runoff auto-correlation. Comparisonof the four models show that models based on auto-correlation are more efficient thanmodels purely based on hydro-climatic teleconnection, and coupling model has higherprecision than time series model.After this, with water demand analysis of Changma irrigation area made, aforecasting-operation coupling model is built based on the rolling-horizon procedure. Areservoir optimal operation model is built taking shortage index as the objectivefunction, and improved dynamic programming algorithm is used to get the solution.Four decision strategies based on different prediction models and different decisionmodels are used to do reservoir operation during drought period of the year2003to2011. Results show that, shortage index is the largest under the actual operation strategy;perfect forecast-optimal operation strategy gets the best result; coupling prediction model-optimal operation strategy is worse than perfect forecast based strategy but betterthan time series model-optimal operation strategy. Therefore, by taking into account ofwater demand in the future and runoff predicted results, measures can be efficientlytaken in prevention of the possible water shortage in the future and get more benefits.On the other hand, runoff predicted results under rolling-horizon procedure during2003-2011show that, in most cases coupling model can get more accurate results thantime series model, but when extreme events happen, both two models provide badresults.