Characteristics Analysis And Predicting Method Of Runoff In Upper Reaches Of Fenhe River

Water resource was one of the most basic elements for supporting the alllives on the earth and realizing sustainable development the change of runoff fa-ctor played a leading role. And it had a profound effect on resource environmentand regional economy development. The study of runoff evolution law was theprerequisite and foundation of rational exploitation and effective utilizationof water resource. Under the combined influence of atmospheric circulation,solar activity, hydrologic meteorological elements, physical geographyand many other uncertain factors, the spatial-temporal evolution of river runoffpresented uncertainty, multiple time scales, randomness, chaos, weak dependent,highly complex nonlinear and non-stationary characteristics. These changecharacteristics put forward new challenges for mid-long term prediction ofrunoff. Many previous studies about runoff prediction were conducted under theassumption that the runoff time series were stationary. However, the runoff timeseries were typical non-stationary series. That assumption often resulted in lowprediction accuracy. According to the non-stationary characteristics of change ofhydrologic element, it would be especially important to introduce new analysismethod and research approach. In this paper, the hybrid prediction model basedon the method-Empirical Mode Decomposition (EMD), which was used to makethe series stationary, was used to the mid-long runoff forecasting. Thehydrologic system was analyzed and predicted from the perspective of variablespatial-temporal scale and non-stationary.Take the annual and monthly runoff data covering1956～2000from4 hydrological stations of the upper reaches of the Fenhe River as an example,first the mathematical statistic methods, hypothesis-testing method, run theory,anomaly theory, long range correlation theory, chaos theory, etc were used tosystematically research the evolution regularity of runoff. On this basis, thehybrid prediction models of auto regressive model (AR), mean generatingfunction-stepwise regression model (MGF-SR), mean generatingfunction-optimum subset regression model (MGF-OSR), Nash nonlinear greybernoulli model (Nash NGBM(1,1)) and chaotic-least square support vectormachine model (C-LSSVM) based on EMD, were established to predict theannual runoff series of the upper reaches of the Fenhe River. The main contentand the main results of this research were as follows:(1) Change characteristic analysis of intra-annual runoff in the upperreaches of the Fenhe RiverIntra-annual runoff distribution of the upper reaches of the Fenhe River wasvery unevenly. The largest runoff was in August, and the minor runoff was inDecember to following February. The change trend of throughout runoffand flood season runoff was of high consistency. Intra-annual runoff of theupper reaches of the Fenhe River presented bimodal type, and the peakrespectively appeared in March and August. As time goes on, the concentrativedegree of intra-annual distribution of runoff showed increasing firstly, thendecreasing, the uniformity decreasing firstly, then increasing, tending to besteady finally.(2) Change characteristics analysis of inter-annual runoff in the upperreaches of the Fenhe RiverThe inter-annual change of runoff was roughly divided into four stages:1956~1970、1970~1994、1994~1996and1996~2000. The runoff in1956~1970increased fluctuantly and tended to be wet water, in1970~1994decreased fluctuantly and tended to be dry water. The runoff of1994~1996presented uptrend, i.e.1994~1996was wet water year. And the runoff of 1996~2000presented downtrend, i.e.1996~2000was dry wet water year.Continuous dry water years was more than continuous wet water years, it wasconsistent with actual situation that Fenhe river basin was prone to drought. Theinter-annual change of upper reaches of the Fenhe River was larger. The largestvariation coefficient and extremes ratio appeared in Shangjingyou station. Thevariation coefficient and extremes ratio of the other three stations were all lessthan this station’. This was mainly because the regulation function of Fenhereservoir on lower reaches.(3) Normality, dry and wet water, stationary, trend and long rangecorrelation characteristics analysis of runoff change in the upper reaches of theFenhe RiverThe annual runoff series of the upper reaches of the Fenhe River showedobviously sharp peaks and fat tail, right-skewed features. And it is typicallynon-normal time series. The number of dry water year was more, the durationlonger, the longest reached to4years. The duration of wet water was shorter,only1~3years. Those showed that the continuous dry water situation moreeasily happened and the annual runoff time series was typically non-stationarytime series. The annual runoff of Fenhe reservoir, Zhaishang and Lancunhydrologic stations presented obvious downtrend. This downtrend became moreand more obvious from upstream to downstream. As well as Shangjingyoustation showed downtrend, but this downtrend was not significant. The annualrunoff of the upper reaches of the Fenhe River had obvious continuity andfractal structure, and this continuity was positively related, i.e. the general trendof future is same with the history. The annual runoff series had10yearsnon-periodic cycle length in the upper reaches of the Fenhe River.(4) Study of hybrid predicting model of annual runoff based on EMDThe annual runoff series of4hydrological stations in the upper reaches of theFenhe River were all non-stationary time series. The EMD method was used tomake the runoff series of4hydrological stations stationary. The results showed after the EMD treatment the several order IMFs of4hydrological stationschanged into stationary seriesAiming at the non-stationary characteristics of runoff time series, the ARmodel was established based on EMD. The results showed the predictingaccuracy of4hydrological stations were all over80%, among them, thepredicting accuracy of Shangjingyou, Zhaishang and Lancun stations all reached100%. The predicting accuracy of hybrid predicting model was obviously higherthan soleAR model’.Aiming at the non-stationary characteristics of runoff time series, theMGF-SR model and MGF-OSR hybrid model based on EMD were established.The results showed the fitting accuracy of MGF-SR model of4hydrologicalstations were between70%and85%, the highest predicting accuracy was only40%. Whereas the fitting accuracy of hybrid model of EMD and MGF-SR werebetween85%and92.5%, and the predicting accuracy were all over60%,moreover, the predicting accuracy of3hydrological stations reached80%. Thefitting and predicting effect of the latter was obviously better than the former.The fitting accuracy of MGF-OSR model of4hydrological stations wasbetween87.5%and90%, the predicting accuracy were all100%, and thedeterministic coefficient were between0.503and0.732. Whereas the fittingaccuracy of hybrid model of EMD and MGF-OSR were all over92.5%, thepredicting accuracy was all over100%, and the deterministic coefficient wasbetween0.975and0.993. The fitting and predicting effect of the latter wasobviously better than the former.Aiming at the non-stationary characteristics of runoff time series, the NashNGBM(1,1) hybrid model based on EMD was established. The results showedthe fitting accuracy of Nash NGBM(1,1) model based on PSO of4hydrologicalstations were between72.5%and82.5%, the most was82.5%, the highestpredicting accuracy was80%, the deterministic coefficient were all less0.9.Whereas the fitting accuracy of Nash NGBM(1,1) model based on EMD and PSO of4hydrological stations were all over90%, the highest was95%, thepredicting accuracy all reached100%, the deterministic coefficient were all over0.98. The fitting and predicting accuracy of the later was obviously more thanthe former.Aiming at the non-stationary characteristics of runoff time series, theC-LSSVM hybrid model based on EMD was established. The results showedseveral order IMFs of Shangjingyou, Fenhe reservoir, Zhaishang and Lancunstation, except for the IMF5of Shangjingyou, Fenhe reservoir and the IMF4ofZhaishang, Lancun station, the14series had chaotic characteristics. The fittingaccuracy of C-LSSVM model of4hydrological stations were between58%and68%, the most was68%, the predicting accuracy of Shangjingyou and Fenhereservoir were100%, and the predicting accuracy of Zhaishang and Lancunwere80%. Whereas the fitting accuracy of C-LSSVM model based on EMD of4hydrological stations were all97%, the predicting accuracy all reached100%.The fitting and predicting accuracy of the later was obviously more than theformer.The hybrid predicting model based on EMD had higher prediction accuracythan sole model, which showed the advantage of making series stationary ofEMD was put forth sufficiently. The hybrid predicting method based on EMDwas effective and feasible to improve the mid-long term runoff predictingaccuracy. This method provided new ways for mid-long term predicting ofrunoff.