| Following the requirement of the reservoir operation section of North-west electricity management bureau in developing a 10 day runoff forecasting scheme and considering the situation of no suitable forecasting model to serve the purpose, this thesis, from Ge shouxi's and WMO's point, defined a group of standards which are practicability, information ability which means the ability of a model to adopt information contend in observed data, extension capacity and robustness, being used for evaluating performance of a forecast model and with them makes a comparison among systematic models as multi-regression model, generalized tank model, neural network method and system model of genesis (in short SMG), and one conceptual model, the Xinganjiang model. After this, a suitable model was proposed to be used to forecast 10 day runoff towards Liujiaxia and Longyangxia reservoirs both of them being settled on the upstream of the Yellow river. The conclusion drawn from the studies are as follows.(1) All systematic models mentioned above are of practicability when there is adequate observed data in rainfall and runoff.(2) No matter what is the data condition, being in plentiful or shortage, the performance of neural network model showed being more or less the same both in calibration and verification stages,which means that neural network model is good at robustness and extension capacity. The forecast result showed that the model has no information ability because it fails to use updating model.(3) SMG model is not only good at practicability but also information ability due to real time updating, which guarantees the precision of the model forecasts. And no matter what stage is, calibration or verification, and what kinds of runoff situation, most plentiful runoff or most shortage, very close forecast precision in each case is obtained. It is suggest that the SMG model be as neural network model in extension capacity. This may be because it comes from deduction of formula with genesis base.(4) Both multi regression model and generalized tank model have high precision in calibration stage, but sharply decrease in verification. Beside that, the quantity of observed data is vital for these two models. So it is suggest that the models be not adopted when data sequence is less than 12 years.(5) Real time updating is quite important for good forecast. However, if there is no new information joining the input of updating model, the neural network updating will not only lose its effective force but also bring even worse results.(6) Dividing catchment into subbasin according to physiographic characteristics of the catchment and inputting information which well represents current situation is a very effective method in increasing forecast precision for all the systematic models mentioned.(7) From analysis, the SMG model with decaying recursion updating is the most suitable model for reservoir 10 day runoff forecast at Liujiaxia and Longyangxia. And neural network model is the next.(8) The Xinanjiang model has been failed to be calibrated in both catchment. There are perhaps three reasons for that. The one is the model being not suitable to the region; the tow is inadequate observed data for calibrating parameters of the model and the three is lack of the experience of the user in use of the model.As to the hydrological characteristics of the basin, the last conclusion comes from that no matter which catchment, Tao or Tangnihai basin, there are significant differences in rainfall and runoff between upstream and downsteam. For Tao river basin, the key runoff area located at Mianxian and upward region, and Maque and upwards for Tangnihai basin.
|
PDF Full Text Request